9HSTFMG*afejbj+
ISBN 978-952-60-5491-9 ISBN 978-952-60-5492-6 (pdf) ISSN-L 1799-4934 ISSN 1799-4934 ISSN 1799-4942 (pdf) Aalto University School of Electrical Engineering Department of Communications and Networking www.aalto.fi
BUSINESS + ECONOMY ART + DESIGN + ARCHITECTURE SCIENCE + TECHNOLOGY CROSSOVER DOCTORAL DISSERTATIONS
Aalto-D
D 211
/2013
Renaud-A
lexandre Pitaval
Coding on F
lag Manifolds for L
imited F
eedback MIM
O System
s A
alto U
nive
rsity
Department of Communications and Networking
Coding on Flag Manifolds for Limited Feedback MIMO Systems
Renaud-Alexandre Pitaval
DOCTORAL DISSERTATIONS
��������� ��������������� �������������������� �����������
����������������� ���� ��������������������������������
�����������������������
������� ������� ����������������� �������� ����������� ����������� ����������������������!�"��������� ��������������������� ��������������#���� ����#��� �!�������������$����������������������� �����������������������������%��� �����&���������'�
����������������� ����!������������"�������"��#���$����!��$$������������� ��%��&��"
��#�������"#��!�����( ��'�)���� **���
� �����������( ��'�)���� **���
�����$��������$�����( ��'������%'�+���!�(� ������� ���!��������'�%� *�+�����!���� ������� � *�!�,����
�##�����( ��'�-��� ��.���� ��*!���*����� ���!�����
��������� ��������������� �������������������� �����������
/�-����0���$�� ��(������
1�23�456047�08�07&4�04�1�23�456047�08�07&4�08�9���:�1��30+��5440&4�&�1��3��5440&4�&�9� ���:�1��3��5440&4&��9���:�����;��� '���-3;1�23;456047�08�07&4�08�
�� ����)��<���*������
,����
�'���������������������(�)�)*��++,,,(-�,,,./�����%%%)�����)!�
��� ������������������ ���!�� �$��!� �������������������������������������� ���� ��������������������������������
��'��� ����"���� �#��������#�����������
����$�%��������� ��������������������&��'�������
�����������(��!�������%�����������������$��)�����$���#�)�)��&��*++,*-+.�
-�����!������� �� ���������)"�����
0�������#���'$�����+/�0����*-+.� �����!� ���!����.+�0�������*-+/�
���$��������#�'��� "������1����2*���%�������*-+.� ���"��"�#����"�
0���"��# �������������������1��$$���3���"������������2
�'������)"��� ��������� ��"��%"��������������������������������������������������������%���
��������%���1����2����"��3�����������������������"���!���������� ��"������������� ���������1���2�����"�������������4���������%���������������5�����������������3�����6�������"�������!���������� ������������������"����"��������������� ���������"����4�)"��������%���������������������������������%���������������������5�'"�����"�������!��������������"��������� �������'���� ������%����������������������"�����'��"��"�������������4�)���������������"�����������"��������������6�����%����������� ���4��������������������'��"���������������"���������!��5��"���%������ ��������������7������������� ���4�8��"��������������!����������'��������5��"���������������������������������������6������� ����� ������ ���������������������������� ���4������"����"����5������������������� ��������������4�����������%���5��������� ����������������� ���������%�����������%���������������� ������ �"�%���%"����4�)"���"������ ������������������"������������ ��"���%����������%��������������������!��������1������������%�������2����!������3�����6�����5����'�������������"���������%�������������������%�������������������4������'�����������������������������5��"��%����������������������������%"�����������4������%��������%"����������������������������!�������� ��������������'��"���"�������'���%��������������%�����4�����������������5��"�������%��� ��������� ��������������%�������!���������4��%���������������5�"�!���������������%�����������%��%�����������������������������������������6�����5������������������������������� ��%����������%�4����������������������������������������������%�������5���%������������������������������������������4����"������������3���������9��������������6���"��7����������������� ������ ���4���!������3�����6������������"�� ���9�����7������������� ��3�����6���������%��%����4�����5��"��%����������� ���������������������������������� �����������������'��"������������!��4�
4��%����������������:�3�����6�����:�������:����5����� �5�7��������5����� ����
�* 1#������2;<=�;>*�?-�>/;+�;� �* 1#�!2;<=�;>*�?-�>/;*�?�
�� ��+<;;�/;./� �� 1#������2+<;;�/;./� �� 1#�!2+<;;�/;/*�
���������!#�'��� ��@������� ���������!#������"@������� 5���*-+.�
��"��+?+� ���"��%A,,���4�,(�&A��B&A;<=�;>*�?-�>/;*�?�
Preface
The research work for this doctoral thesis has been carried out at the
department of Communications and Networking of Aalto University. The
work was funded by the Finnish Funding Agency for Technology and Inno-
vation, the Academy of Finland, Nokia Foundation, and HPY Foundation.
First and foremost, I wish to express my gratitude to my supervisor Prof.
Olav Tirkkonen for his support and guidance. Working with Prof. Tirkko-
nen was a great pleasure and a continuous journey of learning through
many enlightening discussions.
Part of this work was also done at the Image Processing and Commu-
nications Laboratory, Queen’s University, Canada, during 2009-2010. I
would like to sincerely thank Prof. Steven D. Blostein for his kindness
and support during my stay at Queen’s. I wish also to thank Lect. Wei
Dai for welcoming me in his group at Imperial College London for a re-
search visit in Spring 2013.
I am thankful to the co-authors of the papers included in this thesis,
Karol Schober, Prof. Risto Wichman, and Ashivin Srinivasan, and espe-
cially Dr. Helka Määttänen for valuable instructions when I started at
the department. I would like also to thank the thesis preliminary exam-
iners, Prof. David J. Love and Dos. Jyrki Lahtonen, for their constructive
comments; and Prof. Robert Calderbank for accepting to be the defence
opponent.
I am grateful to all nice friends and colleagues I have met during these
past years in Finland, Canada, and UK. Mes remerciements les plus
sincères sont destinés à ma chère et tendre épouse, Viet-Anh, pour son
amour, son soutien, et son illustration plein d’humour de ce manuscrit.
Helsinki, November 28, 2013,
Renaud-Alexandre Pitaval
i
Preface
ii
Contents
Preface i
Contents iii
List of Publications v
1. Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objective and Scope . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Contribution and Structure of the Thesis . . . . . . . . . . . 3
1.4 Summary of Publications . . . . . . . . . . . . . . . . . . . . . 5
2. System Model 9
2.1 Signal and Channel Model . . . . . . . . . . . . . . . . . . . . 9
2.2 Base Station Cooperation . . . . . . . . . . . . . . . . . . . . 10
2.3 Codebook-Based Precoding . . . . . . . . . . . . . . . . . . . . 11
2.4 Low-Complexity Constraints . . . . . . . . . . . . . . . . . . . 11
3. Flag Manifolds: Code Designs and Geometry 13
3.1 Flag Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Spherical Embeddings and Chordal Distances . . . . . . . . 15
3.3 Codebook Designs . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4 Lloyd’s Algorithm and Centroids . . . . . . . . . . . . . . . . 17
3.5 Density and Packing Bounds . . . . . . . . . . . . . . . . . . 18
4. Grassmannian Pakings for 2-Tx MIMO 23
4.1 Grassmannian Codebooks from Spherical Arrangements . . 23
4.2 Low Implementation-Complexity Codes . . . . . . . . . . . . 26
4.3 Quantization Error Analysis . . . . . . . . . . . . . . . . . . . 28
5. Flag and Stiefel Orbit Codes 31
iii
Contents
5.1 Orbits of Projective Group Representations . . . . . . . . . . 31
5.2 Extraspecial Group Code Constructions . . . . . . . . . . . . 35
5.3 Expansion to Stiefel Codes . . . . . . . . . . . . . . . . . . . . 38
6. Joint Grassmann-Stiefel Codes for Product Codebooks 41
6.1 Product Codebook-Based Precoding . . . . . . . . . . . . . . . 41
6.2 Joint Grassmann-Stiefel Codebooks . . . . . . . . . . . . . . 43
6.3 Lloyd-type Algorithm for Joint Grassmann-Stiefel Codebook 46
6.4 Codeword Selections . . . . . . . . . . . . . . . . . . . . . . . 47
7. Flag Codebooks for MIMO Systems with Linear Receiver 51
7.1 Achievable Information Rates . . . . . . . . . . . . . . . . . . 52
7.2 Linear Receiver versus ML Receiver . . . . . . . . . . . . . . 52
7.3 Codebook Designs . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.4 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
8. Conclusions 57
Bibliography 61
Errata 67
Publications 71
iv
List of Publications
This thesis consists of an overview and of the following publications which
are referred to in the text by their Roman numerals.
I R.-A. Pitaval, O. Tirkkonen and S. D. Blostein. Density and Bounds
for Grassmannian Codes with Chordal Distance. In Proceedings of the
IEEE International Symposium on Information Theory (ISIT), Saint Pe-
tersburg, Russia, pp. 2298-2302, Aug. 2011.
II R.-A. Pitaval and O. Tirkkonen. Volume of Ball and Hamming-type
Bounds for Stiefel Manifold with Euclidean Distance. In Proceedings of
the 46th Annual Asilomar Conference on Signals, Systems, and Comput-
ers (ACSSC), Pacific Grove, California, pp. 483-487, Nov. 2012.
III R.-A. Pitaval, H.-L. Määttänen, K. Schober, O. Tirkkonen, and R.
Wichman. Beamforming Codebooks for Two Transmit Antenna Systems
based on Optimum Grassmannian Packings. IEEE Transactions on In-
formation Theory, vol.57, no.10, pp. 6591-6602, Oct. 2011.
IV R.-A. Pitaval, O. Tirkkonen and S. D. Blostein. Low Complexity MIMO
Precoding Codebooks from Orthoplex Packings. In Proceedings of the
IEEE International Conference on Communications (ICC), Kyoto, Japan,
pp. 1-5, June 2011.
V R.-A. Pitaval and O. Tirkkonen. Grassmannian Packings from Orbits
of Projective Group Representations. In Proceedings of the 46th Annual
Asilomar Conference on Signals, Systems, and Computers (ACSSC), Pa-
v
List of Publications
cific Grove, California, pp. 478-482, Nov. 2012.
VI R.-A. Pitaval and O. Tirkkonen. Flag Orbit Codes and Their Expan-
sion to Stiefel Codes. In Proceedings of the IEEE Information Theory
Workshop (ITW), Seville, Spain, pp. 360-364, Sep. 2013.
VII R.-A. Pitaval and O. Tirkkonen. Joint Grassmann-Stiefel Quantiza-
tion for MIMO Product Codebooks. Accepted for publication in IEEE
Transactions on Wireless Communications, 13 pp., Oct. 2013.
VIII R.-A. Pitaval, A. Srinivasan and O. Tirkkonen. Codebooks in Flag
Manifolds for Limited Feedback MIMO Precoding. In Proceedings of
the 9th International ITG Conference on Systems, Communications and
Coding (SCC), Munich, Germany, pp. 1-5, Jan. 2013.
Author’s Contribution
The author has had a leading role in writing all the publications. The
author has had a leading role in the planning and the analysis of Publi-
cations [I, II, IV, VII]. The author has participated in theoretical mod-
eling and design of experiments in Publications [III, V, VI, VIII]. The
author has conducted the performance simulations for all publications ex-
cept Publication VIII.
vi
List of abbreviations and symbols
Abbreviations
2D two dimension
3D three dimension
4D four dimension
4G 4th generation systems
BS base station
CB codebook
CoMP coordinated multi-points
CSI channel state information
dB decibel
EP Equal transmit-power per-antenna
FDD frequency-division duplexing
HB Hamming bound
i.i.d independent and identically distributed
LTE long-term evolution
MIMO multiple-input multiple-output
ML maximum likelihood
MMSE minimum mean square error
MS-MUB maximal set of mutually unbiased bases
MU multi-user
MUB mutually unbiased bases
pdf probability density function
SISO single-input single-output
SNR signal-to-noise ratio
SU single-user
Rx receive antenna
ZF zero forcing
vii
List of abbreviations and symbols
Symbols
C complex field
CP1 complex projective line
C codebook
Ck border of the kth spherical cap
ck kth spherical codeword
Ck kth codeword
Cpr product codebook
dc chordal distance
df flag chordal distance
dg Grassmann chordal distance
dmu MUness flag distance
dp permutation-invariant flag chordal distance
ds Stiefel chordal distance
D dimension of embedding sphere
DM(C) average squared distortion of codebook C on manifold MD diagonal matrix
det determinant
E[·] mathematical expectation
F part of linear receiver
fd2g pdf of squared Grassmann distortion
F lCnt;s1,...,sr nt-by-ns complex flag manifold with equivalence under r
unitary groups of size si
�F lC
nt;s1,...,sr nt-by-ns complex flag manifold with equivalence under r
unitary groups of size si and permutation of the r groups of
substreams
FCnt,ns
nt-by-ns complex flag manifold with equivalence under ns
phase rotation�FCnt,s1,...,sr nt-by-ns complex flag manifold with equivalence under ns
phase rotation and permutation of the ns columns
g an abstract group element
G an abstract group
G large-scale path gain matrix
GCnt,ns
complex nt-by-ns Grassmann manifold
H fading channel
Heff effective fading channel
viii
List of abbreviations and symbols
Hk fading channel from kth BS
Hls large-scale fading channel
Hss small-scale fading channel
I average information rate
Ilr information rate with linear receiver
Iml information rate with ML receiver
Inm maximum achievable rate without transmission rank con-
straint
Ins maximum achievable rate with ns streams
Int nt-by-nt identity matrix
Int,ns nt-by-ns identity matrix
Kb number of bits to represent a real number
log2 binary logarithm
M manifold
n noise vector
nbs number of BSs
ncb size of codebook
nr number of receive antennas
ns number of streams
nt number of transmit antennas
Ng Grassmann codebook cardinality
Ns Stiefel codebook cardinality
P permutation matrix
PUnt projective unitary group of size nt
qC abstract quantization map associated with Cq∗ optimum quantization map using information rate
qc quantization map using chordal distance
qg quantization map using Grassmann chordal distance
r number of subgroups of streams
R radius of embedding sphere
Rk kth Voronoi cell
R real field
sk kth group of substreams
Sr symmetric group of degree r
SD(R) D-sphere of radius R
SOnt special orthogonal group of size nt
u(ns) vector space of skew-Hermitian ns-by-ns matrices, Lie al-
gebra for Uns
ix
List of abbreviations and symbols
Unt unitary group of size nt
U unitary matrix
Ui unitary matrix of size si
VCnt,ns
complex nt-by-ns Stiefel manifold
V random source on Stiefel manifold
v point on sphere
vk kth column of V
V nt-by-ns Stiefel matrix
Vk nt-by-sk Stiefel submatrix of V
Vls ns-largest right singular vectors of Wls
Vss ns-largest right singular vectors of Wss
w precoding vector
W precoding matrix
Wls large-scale precoder
Wss small-scale precoder
x transmitted symbol vector
y received signal
yk kth column of Y
Y nt-by-ns Stiefel matrix
Yk nt-by-sk Stiefel submatrix of Y
Z(G) center of group G
Zk centroid of kth Voronoi cell
αk large-scale path gain from kth BS
δM(C) minimum chordal distance of codebook C on manifold Mδg minimum Grassmann chordal distance
δs minimum Stiefel chordal distance
γ snr per-stream
γk post-processing SINR of the kth data stream
λk kth singular value of Heff
θ angle
φ angle
ϕ angle
μ(B(rn)) normalized volume of a ball of radius rn
σ permutation
σk kth singular value of H
·H Hermitian transpose
x
1. Introduction
1.1 Background
Multi-input multi-output (MIMO) techniques are key technologies to en-
hance spectrum efficiency of wireless systems. Performance heavily de-
pends on the channel state information (CSI) available at the transmitter.
MIMO using linear precoding have been shown to achieve large capac-
ity gains over traditional single-input-single-output (SISO) systems [32,
87]. Linear precoding for single and multiple stream transmission, a.k.a.
beamforming and multiple beamforming, has been intensely investigated
for point-to-point communications [75, 89]. Without channel reciprocity,
e.g. in frequency-division duplex systems, the only way to acquire CSI
is through a limited feedback channel. Fortunately, few bits is usually
enough to fill most of the gap between open-loop and closed-loop capac-
ity [55]. A widely applied method is to use codebook-based precoding in
which the receiver selects a precoding codeword from a predefined code-
book and feeds back the index to the transmitter. Since it is more impor-
tant to feed back the channel direction than the channel beam gain [23],
the quantization of the channel is often done with a rectangular unitary
code.
In point-to-point communications with maximum likelihood receiver, the
performance of a unitary precoding codebook is related to its interpreta-
tion as a discretization of the Grassmann manifold [22, 23, 53, 55, 56, 61].
Codebook design criteria are based on extremization of average distortion
metrics [22, 59, 63, 71, 72, 95]. The information rate is approximately a
function of the distortion rate of the codebook associated with Grassmann
chordal distance as a quantization map [22,23]. In [56,61] the beamform-
ing codebook design problem for uncorrelated channel was linked to a
1
Introduction
suboptimal approach, the Grassmannian line packing problem, i.e. max-
imizing the minimum distance of the codebook. Extension to correlated
channel through codebook rotation is then provided in [54,95].
Grassmann packing, originally a mathematical problem of independent
interest [18], was thus retained as a method for codebook design. Ex-
tensive tables of real Grassmannian codes could be found in the liter-
ature [83]. For the complex Grassmannian, fewer results were avail-
able [15, 79, 85]. Complex Grassmann code construction was hence ad-
dressed in [24, 56, 64, 72, 76, 95, 96], codes being generated by computer
searched by either directly minimizing the distortion of the codebook us-
ing vector quantization algorithms such as Lloyd-type algorithms [64,72,
95]; or maximizing its minimum distance with brute-force search [56],
modified Lloyd’s algorithm [96], alternating projection algorithm [24], and
expansion-compression algorithm [76]. Analytical constructions were pre-
sented in [2, 4–6, 96] with application also to non-coherent MIMO space-
time coding [97], and low-complexity implementation codebooks for pre-
coding [44,51,60].
While Grassmann precoding has attracted much attention, other trans-
mission scenarios or constraints may lead to the need of designing code-
books in other spaces. Practical codebooks in industry standards have
been designed according to power and implementation-complexity con-
straints [51]. With an MMSE or ZF receiver, the set of non-equivalent pre-
coding matrices is not a Grassmann manifold anymore [57]. In [57], the
proper space of quantization is presented along with some optimum ana-
lytical packings for 2 transmit antenna systems. For bit-interleaved mul-
tiple beamforming, it is shown in [77] that the optimum precoder is only
invariant under a diagonal unitary transform, a corresponding Lloyd’s al-
gorithm is presented.
Further development of MIMO techniques towards network-level pro-
cessing are expected to bring performance enhancements. In MIMO broad-
casting, unitary codebook-based precoding has been well-investigated in
a multi-user (MU) setting with one-stream transmission per-user [43,47,
90,91]. In [47], codebook design on a different Riemannian manifold than
the Grassmann manifold is considered, and accordingly, a systematic con-
struction of structured codebook is presented. While considering different
system models, the manifolds discussed in [47] and [57] are the same,
and the proposed distance metrics are equivalent up to a scaling factor.
Similar equivalence classes are described in [90, 91]. To deal with the
2
Introduction
specific features of base stations cooperation or coordinated multi-points
(CoMP) [46, 48, 88], a product codebook quantization strategy was pro-
posed in [16].
Interpreting all those designs in a unified manner, the spaces of dis-
cretization to consider are quotient spaces of unitary groups, and although
not recognized as such, so-called flag manifolds. Literature on flag mani-
folds can be found in quantum theory context, e.g. in [62,68,98].
1.2 Objective and Scope
The goal of the thesis is to design low-implementation complexity precod-
ing codebooks with analytic and algebraic methods. Finding optimal pre-
coding codebooks is equivalent to the problem of discretizing a manifold.
Manifold discretization is a generic mathematical problem covering for
example the particular case of spherical arrangements or real Grassman-
nian discretization which have been addressed due to their relevance to
many fields of science such as chemistry, biology and physics. While there
is an increased interest in manifold discretization from communications
engineers, there are few results from the mathematical literature. An-
alytical discretization of complex manifolds yields mathematically struc-
tured codebooks which are amenable for implementation. In this work,
we concentrate on discretization techniques for quotient spaces of unitary
groups, so-called flag manifolds, including Stiefel and Grassmann mani-
folds.
1.3 Contribution and Structure of the Thesis
This dissertation contributes to the field of coding theory for unitary man-
ifolds and codebook design for MIMO precoding.
We address several codebook constructions in the literature as a unified
mathematical problem of discretization of flag manifolds. The Conway-
Hardin-Sloane spherical embedding of the Grassmann manifold with chor-
dal distance is generalized to other flag manifolds, and modified for equal-
power per-antenna codebooks. The spherical embeddings allow leverag-
ing results from the mathematical literature. For Lloyd’s algorithms, we
leverage a result on centroid computation on Euclidean embedded sur-
faces applicable to manifolds equipped with a chordal distance. In par-
3
Introduction
ticular, we derive a closed-form centroid computation in the Stiefel mani-
fold. The spherical embeddings induce also coding bounds from the sphere
packing literature.
A fundamental problem of coding theory is to establish the maximum
cardinality of a code for a given distance. The well-known Hamming
bound partially answers this question. We have investigated the density
of packings in the complex Grassmann and Stiefel manifolds equipped
with chordal distance. We have computed the exact volume of the Grass-
mann and Stiefel manifold induced by their chordal distance. This has
application in the evaluation of the volume of a small metric ball which is
critical to derive Hamming-type bounds. We provide a refinement of the
Hamming bound for Grassmannian codes and a generalization of a bound
on minimum distance previously proven only for line packings. This result
is later generalized to all manifolds with a metric induced by an embed-
ding in a Euclidean hypersphere. For these manifolds, of which the Stiefel
manifold, this provides results generalizing previously known bounds on
codes in the unitary group.
We then investigate explicit analytical code constructions, looking for
codebooks having their entries in a limited set of complex numbers. For
the special case of two antenna precoding codebooks, the 2x1 complex
Grassmannian is isometric to a real sphere, so that designing 2x1 complex
Grassmannian packing is equivalent to the real sphere packing problem.
Based on the extensive literature on this topic, we have derived optimal
closed-form codebooks using a simple isomorphism. Using the simple ge-
ometry of some of these codebooks, we also derived closed-form expres-
sion of the corresponding SNR gain due to beamforming. Additionally,
we investigate codebooks based on other spherical arrangements, such as
solutions maximizing the harmonic mean of the mutual distances among
the codewords. We found that in most of cases, Grassmannian codebooks
based on these other spherical arrangements outperform codebooks from
Grassmannian packings. The reason is that the mean distance provides a
better approximation of the average distortion than the minimal distance.
Next, to address the problem of manifold discretization in more general
manner we use the concept of an orbit of a symmetry group. We general-
ize the concept of spherical orbit codes to flag orbit codes and derive ba-
sic properties. For flag codes, projective unitary representations of finite
group are of specific interests. We consider some finite groups having ap-
propriate representations and find appropriate initial points heuristically.
4
Introduction
We give explicit codes in Grassmann and other simple flag manifolds in
2D and 4D. Using a construction related to representation theory, fami-
lies of packings are obtained for any dimension of a power of a prime. We
prove that some of these packings are optimal in relation with a power
per-antenna constraint. We also investigate Stiefel orbit codes arising
from the linear representation of the projective group considered. By do-
ing so, one obtains expansions of the Grassmann orbit codes to the Stiefel
manifold.
Then, we discuss codebook designs for two different transmission sce-
narios. For a base station cooperation system, we consider a product
codebook strategy where a single small codebook is implemented at the
receiver. We show that near optimal performance can be reached with
an appropriate choice of Stiefel representatives of Grassmann codes. Ac-
cordingly, we propose a novel joint Grassmann-Stiefel codebook design
aiming at good quantization of Grassmann and Stiefel manifolds with a
single codebook. A Lloyd-type algorithm generating a Stiefel codebook
conditioned on a fixed Grassmannian codebook is presented. Further-
more, concrete examples of analytical joint Grassmann-Stiefel packings
are also given. We also discuss low-complexity suboptimal codeword se-
lection methods.
Finally, the codebook design problem for MIMO with a linear receiver
is related to a discretization problem of generalized flag manifolds. With
a linear receiver, we show that the spaces of equivalent precoders are
simple permutation-invariant flag manifolds. We describe a Lloyd-type
algorithm for these spaces and compare the achievable rate with the gen-
erated codebooks by simulation. When the number of streams, and the
number of receive and transmit antennas are the same, the simulations
show that the marginal gain from precoding is relatively small with few
feedback bits for more than two transmit antennas. This differs from
low-rank transmission with optimum receivers, where a small number of
feedback bits allows near-optimal channel adaptation.
1.4 Summary of Publications
This thesis consists of an introductory part and eight original publica-
tions.
The first two papers are discussing coding theoretical problems on the
Grassmann and Stiefel manifolds.
5
Introduction
In Publication I, the density of Grassmann codes with the chordal dis-
tance is investigated. From the observation that the kissing radius can-
not be determined solely from the minimum distance, upper and lower
bounds are provided, along with the corresponding bounds on the den-
sity. This leads to a refinement of the Hamming bound for Grassmannian
codes. Finally, we provide explicit bounds on code cardinality and mini-
mum distance, notably a generalization of a bound on minimum distance
previously proven only for line packings.
In Publication II, the density of Stiefel codes is investigated. We com-
pute the volume of the Stiefel manifold induced by the chordal distance.
This has a direct application for evaluating the volume of a small metric
ball critical to derive Hamming-type bounds. Using a spherical embed-
ding argument, we provide results generalizing previously known bounds
on codes in the Grassmann manifold and the unitary group.
The following four papers deal with explicit code constructions.
In Publication III, we construct Grassmann beamforming codebooks for
two transmit antenna systems. Using an isometry, we show that the dis-
cretization problems are directly solved by corresponding spherical codes.
Notably, the Grassmannian line packing problem is equivalent to the
Tammes problem on the real sphere, so that optimum spherical packings
give optimum Grassmannian packings. Moreover, a simple isomorphism
enables to analytically derive simple codebooks in closed-form having low
implementation complexity. Using the simple geometry of some of these
codebooks, we derive closed-form expressions of the probability density
function (pdf) of the squared quantization error. We also investigate code-
books based on other spherical arrangements, such as solutions maximiz-
ing the harmonic mean of the mutual distances among the codewords,
which is known as the Thomson problem. We find that in most of the
cases, Grassmannian codebooks based on these other spherical arrange-
ments outperform codebooks from Grassmannian packing.
In Publication IV, a construction of Grassmannian packings related to
representation theory is applied to build implementation-friendly MIMO
precoding codebooks when the number of transmit antennas is a power
of a prime number. Using chordal distance as a metric, some of the cor-
responding packings are optimal by meeting the orthoplex bound. In ad-
dition, by using only some of the codewords, smaller packings satisfying
an equal-power per antenna constraint can be constructed. Optimality
with reference to this constraint is shown by modifying Conway-Hardin-
6
Introduction
Sloane’s spherical embedding of the Grassmann manifold for equal-power
per-antenna codebooks.
In Publications V and VI, we discuss group orbits to construct codes in
Grassmann and flag manifolds, respectively.
In Publications V, to generate Grassmann orbit codes, we look for projec-
tive unitary representations of finite groups. Following this principle, we
derive basic properties and describe explicit constructions of group orbits
leading to some optimum packings in 2 and 4 dimensions.
In Publication VI, we define distances that embed flag manifolds into
Euclidean hyperspheres, providing a generalization of the spherical em-
bedding of Grassmann manifolds equipped with the so-called chordal dis-
tance. For code construction, the center of a finite unitary group has no
effect, and thus it is sufficient to consider its inner automorphism group.
Accordingly, some explicit constructions from projective unitary represen-
tations of finite groups in 2 and 4 dimensions are described. We also give
examples of codes on the Stiefel manifold constructed as orbits of the lin-
ear representation of the projective groups, leading to codes that are ex-
pansions of the flag codes considered.
The last two papers investigate scenarios where Grassmann codebook
design is not optimal, and a larger manifold has to be considered.
In Publication VII, we focus on product codebook design with application
to cooperative MIMO transmission. We show that the Stiefel representa-
tives which are used to realize a Grassmann codebook impact the perfor-
mance of product codebooks. We propose a novel joint Grassmann-Stiefel
codebook design aiming at good discretization of Grassmann and Stiefel
manifolds with a single codebook. The resulting product codebooks show
performance comparable with global Grassmann quantization. To find
low-distortion codebooks, we present a vector quantizer to generate Stiefel
codebooks conditioned on a fixed Grassmann codebook. For this purpose,
we provide an exact solution for computing centroids in the Stiefel mani-
fold with chordal distance. Furthermore, concrete examples of analytical
joint Grassmann-Stiefel packings are given. Additionnally, we discuss
low-complexity codeword selection methods.
Publication VIII interpretes unitary codebook design problems for var-
ious precoded MIMO scenarios as generalized discretization problems of
flag manifolds. As a concrete example, we consider codebooks for MIMO
transmission with linear receivers. In this case, the problem reduces
to discretizing permutation-invariant flag manifolds. A corresponding
7
Introduction
Lloyd’s algorithm is given, providing low-distortion codebooks. We found
that a full-rank MIMO system with the same number of transmit and re-
ceive antennas, the gain of precoding with linear receiver is small. This
differs from the behavior of low-rank transmissions, where it is known
that a small number of feedback bits allows near-optimal channel adap-
tation.
8
2. System Model
This chapter presents the system model of codebook-based MIMO pre-
coding [51, 55]. For point-to-point MIMO, we consider a uncorrelated flat
Rayleigh MIMO channel. Extension to base station cooperation is mod-
eled by integrating large-scale path gain imbalance between the trans-
mitters, similarly than in [16]. The system model motivates the prime
interest of the thesis which is on code design and construction.
2.1 Signal and Channel Model
Consider a MIMO system with nt transmit and nr receive antennas. After
unitary precoding with W ∈ Cnt×ns , a vector x ∈ Cns×1 of ns ≤ min(nt, nr)
multiplexed streams is transmitted through a fading channel H ∈ Cnr×nt .
The received signal is
y = HWx+ n, (2.1)
= Heffx+ n, (2.2)
n ∈ Cnr×1 denote the noise, and Heff = HW the effective channel. We
assume that the transmitted signal and the noise are Gaussian with co-
variances E[xxH
]= γIns and E
[nnH
]= Inr , where γ is the per-stream
SNR. We assume the entries of the channel H are independent and identi-
cally distributed (i.i.d) complex normal variables with zero mean and unit
variance.
We concentrate on the properties of W related to steering the trans-
mitted energy to the signal subspace of the receiver and increasing the
capacity. Here, power allocation is considered to be out of scope of precod-
ing and instead form a part of the design of x. We thus have the following
contraints on W: the total transmit power is ns, which is equally shared
among the symbols such that WHW = Ins . The precoder, which is a func-
9
System Model
tion of the received feedback bits, is used for channel adaptation in order
to increase to maximum achievable information rate of the system given
by
I = E[log2 det
(I+ γHH
effHeff
)]. (2.3)
2.2 Base Station Cooperation
In Chapter 6, we consider cooperative transmission from several base sta-
tions to the same user. Then, we define a (nbs × nt) × ns MIMO system
as nbs base stations (BSs) each equipped with nt antennas transmitting
cooperatively ns-data streams. It is assumed that the BSs are able to
instantaneously share the feedback information, e.g. via high speed back-
hauls. The effective channel becomes
Heff = HlsWls, (2.4)
where Wls is an nbsnt × ns aggregate precoding matrix and
Hls = [α1H1, . . . , αnbsHnbs
] = HssG (2.5)
is the aggregate channel matrix where the channels from the BSs to the
receiver are concatenated, and large scale path losses are explicitly taken
into account. The average path gain from the ith BS to the receiver is αi,
incorporating distance-dependent path loss and shadowing. Small-scale
path gains are characterized by the matrices Hi ∈ Cnr×nt whose entries
are assumed to be i.i.d. flat Rayleigh distributed with unit variance. The
aggregate small-scale path gain matrix is denoted Hss = [H1, . . . ,Hnbs]
and the large scale path gains by G = diag(α1Int , . . . , αnbsInt). For α1 =
· · · = αnbs, the model reduces to a classical nbsnt × ns i.i.d. point-to-point
MIMO system. We denote by Vls and Vss ∈ VCnbsnt,ns
the left singular
vectors associated with the ns-largest singular values of HHls and HH
ss , re-
spectively.
It is assumed that the BSs know the large scale path gains of the chan-
nels contained in G and the precoder is constructed in two steps. The BSs
first construct a small-scale precoding matrix Wss. Then the large-scale
precoding matrix applied is Wls =√ntnbs
‖G‖ GWss following the principle of
adaptive precoding for correlated MIMO [54,95]. The normalization guar-
antees that the total transmit power per-stream is one.
10
System Model
2.3 Codebook-Based Precoding
The channel coefficients are assumed to be perfectly known at the receiver
and unknown at the transmitters. We assume that the BS has access to
CSI only through an error-free, zero delay, and limited feedback chan-
nel. The precoding matrix W ∈ Cnt×ns is designed for channel adapta-
tion according to the information fed back by the receiver. To acquire
CSI through a limited feedback channel, we consider codebook-based pre-
coding where the receiver and transmitter share a predefined codebook
C = {C1, . . . ,Cncb}. Elements in the codebook are Stiefel matrices, i.e.
CHi Ci = I ∀i, which are used to quantize the eigendirections of the signal
subspace at the receiver.
The receiver selects a codeword following a quantization rule that ap-
proximates the channel by using the codebook,
qC : {H} → {i : 1 ≤ i ≤ ncb}, (2.6)
and feeds back the index k = qC(H) to the transmitter. Then, the trans-
mitter constructs a precoding matrix based on the CSI received, i.e. W is
a function of Ci. Especially for point-to-point MIMO, a simplified version
is to assume that the transmitter directly picks the precoding matrix from
the shared codebook: W = Ck.
In this setting, there is typically an infinity of precoding vectors lead-
ing to the same performance, and the possible precoding vectors can be
grouped into equivalence classes. It follows that the set of equivalence
classes of precoding matrices is a quotient space of the Stiefel manifold.
In order to have non-equivalent codewords, the codebook should be de-
signed as a discretization of this space of equivalence classes of precoding
matrices. A Stiefel precoding codebook is then generated by taking any
representative in the equivalence classes.
2.4 Low-Complexity Constraints
Practical codebooks in industry standards have been designed according
to additional constraints of interest. In our code construction, we try to
address those constraints. Low implementation-complexity codebooks are
typically characterized through three design constraints [51]:
1. Equal-transmit-power per-antenna (EP). The antennas are used in a
11
System Model
power balanced manner and their average transmit power is kept at the
same level. This constrains the rows in the precoding matrix to have
squared norm equal to ns/nt.
2. Constrained alphabet. The alphabet of the codebook elements is re-
stricted to a small finite alphabet in order to limit the number of mul-
tiplications and storage requirements at the receiver. Often, the alpha-
bet is limited to the unit circle, which is also known for beamforming
as equal gain transmission [52, 63]. This guaranties the EP constraint
above. Restricting the alphabet to {1, −1, i, −i}, i =√−1 further re-
duces complexity by enabling matrix multiplications to be performed
only by conjugations and additions [44,60].
3. Nested property. A lower rank precoding matrix is a submatrix of a
higher rank precoding matrix.
In the current 3GPP LTE-Advanced industry standard, the number of
antennas at the base station may be two, four, or eight. The 2Tx code-
book can be seen as an example of extraspecial group code constructions
described in Chapter 5, cf. codebook B in Table I from Publication IV. For
4Tx, the LTE codebooks were designed based on Householder reflections.
The resulting codebooks have similar distance properties as the codebooks
of Publication IV (codebook C in Table 5.4) and thus perform equivalently
in term of spectral efficiency. As discussed in [1] the implementation ben-
efit from the Householder construction was accidental and valid only for
4Tx; it is also noted that discrete Fourier transform (DFT) codebooks fail
to be a valid design for 8Tx systems. To overcome this issue, the 8Tx code-
book in LTE-Advanced is constructed by concatenating the same 4TX DFT
codeword twice and using a cophasing factor [80]. The resulting codebook
has a structure resembling the product codebooks considered in Publica-
tion VII.
12
3. Flag Manifolds: Code Designs andGeometry
The codebooks addressed have orthonormal columns, and can be inter-
preted as elements in a flag manifold with representative in a Stiefel
manifold. A manifold is roughly speaking a curved space which locally
resembles the Euclidean space. To investigate coding problems, we need
to define a notion of distance. Depending of the application and conve-
nience, several non-equivalent distances has been considered on these
spaces [28]. We focus on chordal distances corresponding to natural Eu-
clidean distances from spherical embeddings. There are two motivations
for this choice. First, we are interested in computable distance functions
consistent with the rectangular unitary matrix representation. Second,
the considered chordal distances for Stiefel, Grassmann, and some simple
flag manifolds have been shown to be related to performance of space-time
codes and precoding codebooks in SU and MU-MIMO [22,35,47,56]. This
choice of distances enforces treatment of the manifolds as subsurfaces of
hyperspheres, and implies that flag codes are a subclass of spherical codes.
3.1 Flag Manifolds
The complex Stiefel manifold VCnt,ns
is defined as the space of orthonormal
rectangular matrices (with ns ≤ nt),
VCnt,ns
={Y ∈ Cnt×ns | YHY = Ins
}. (3.1)
The unitary group Unt = VCnt,nt
is a specific case of Stiefel manifold.
The flag manifold [11,29]
F lCnt;s1,...,sr = VCnt,ns
/(Us1 × · · · × Usr) (3.2)
where ns =∑r
i=1 si, is the set of equivalence class of nt-by-ns Stiefel
matrices where two matrices V,Y are equivalent if there exists a se-
13
Flag Manifolds: Code Designs and Geometry
W W
U1
U2U3
nt
ns s1 s2 s3 ns
ns
(a) Unitary-rotations equivalence (b) Permutation equivalence
Figure 3.1. Illutration of the equivalence relationship considered.
quence of unitary matrices (U1, . . . ,Ur) ∈ (Us1 × · · · × Usr) exist such that
V = Y diag(U1, . . . ,Ur).
The case r = 1 and s1 = ns defines the Grassmann manifold
GCnt,ns
� F lCnt;ns= VC
nt,ns/Uns , (3.3)
which is isomorphically the set of all ns-dimensional subspaces of Cnt .
According to (3.2), the Stiefel manifold VCnt,ns
is not strictly speaking a
flag manifold, but it is also a homogeneous space of the unitary group
VCnt,ns
∼= Unt/Unt−ns and can be seen as an extreme case in (3.2) by setting
r = 1 and s1 = 0.
Another case of interest is s1 = . . . = sr = 1, for which we use the
notation
FCnt,ns
� F lCn;1, . . . , 1︸ ︷︷ ︸
ns
= VCnt,ns
/(U1)ns . (3.4)
We finally consider an additional equivalence relationship in F lCn;s1,...,sr
when sk = ns/r for all k. Given V,Y ∈ VCnt,ns
, one may group their
columns in sub-matrices as V = (V1,V2, . . . ,Vr) and Y = (Y1,Y2, . . . ,Yr),
such that Vk,Yk ∈ VCnt,sk
. Then, we consider them equivalent if they
only differ by a permutation of their r-submatrices, i.e. there exists a
permutation σ such that V = (Yσ(1),Yσ(2), . . . ,Yσ(r)). The permutation
corresponds to an orientation-invariance of the elements. We denote the
corresponding space by
�F lC
nt;s1,...,sr � F lCnt;s1,...,sr/Sr (3.5)
where Sr is the symmetric group whose elements are all the permutations
of the r symbols.
Similarly for the specific case s1 = . . . = sr = 1, we defined
�FCnt,ns
� FCnt,ns
/Sns (3.6)
14
Flag Manifolds: Code Designs and Geometry
3.2 Spherical Embeddings and Chordal Distances
We treat the manifolds of interest as submanifolds of hyperspheres. The
spherical embeddings of the Stiefel manifold and the flag manifolds are
of different natures. The Stiefel manifold has a canonical spherical em-
bedding from the vector representation of rectangular unitary matrices.
The spherical embeddings of the Grassmann manifolds with correspond-
ing chordal distance is obtained from a projector representation [18]. In
Publication VI, we show that all flag manifolds can be understood as sub-
manifolds of the same sphere. Roughly speaking, a (n2t − 2)-dimensional
hypersphere can be decomposed, except for a zero-measure set, so that it
consists of a “fibration” of flag manifolds FCnt,nt
over a (nt−2)-dimensional
hypersphere with some singular submanifolds removed. The remaining of
the sphere, the singularities, correspond to other flags F lCnt;s1,···sr which in-
cludes the Grassmann manifolds GCnt,ns
as special cases. To generalize the
notion of Grassmann distance to more general flag manifolds, we embed
them into a direct product of Grassmann manifolds and take the corre-
sponding chordal distance. This results in embedding into a larger space
than the (n2t − 2)-sphere. Finally, inspired by the literature in quantum
information science on mutually unbiased bases (MUB) [27], we consider
an alternative distance on simple permutation-invariant flag manifolds.
We have the following isometric embeddings
(M, dc) ↪→ SD−1(R) (3.7)
where SD−1(R) ⊂ RD is a (D − 1)-sphere of radius R, with
M dc D R2
VCnt,ns
ds 2ntns ns
GCnt,ns
dg n2t − 1 ns(nt−ns)
2nt
F lCnt;s1,...,sr df r(n2t − 1)
nsnt−∑r
i=1 s2i
2nt
�F lC
nt;s1,...,sr dp r(n2t − 1)
nsnt−∑r
i=1 s2i
2nt
�FCnt,ns
dmu
(n2t2
)− 1(ns−1)(n2
t−ns)
n2t−1
and the chordal distances, defined for Y, Z ∈ VCnt,ns
, representatives of
15
Flag Manifolds: Code Designs and Geometry
their respective equivalence class, are
ds(V,Y) = ‖V −Y‖F , (3.8)
dg(V,Y) =1√2‖VVH −YYH‖F , (3.9)
df (V,Y) =
√√√√ r∑i=1
d2g(Vi,Yi), (3.10)
dp(V,Y) = minP∈Sp
df (V,YP), (3.11)
dmu(V,Y) =
√√√√ns −ns∑
i,j=1
|vHi yj |4. (3.12)
(3.13)
The decompositions of the matrices in column-blocks are given by V =
(V1, . . . ,Vr) = (v1, . . . ,vns) and Y = (Y1, . . . ,Yr) = (y1, . . . ,yns). Here
df is the metric on the flag manifold interpreted as a sequence of Gras-
mannians, and dp is the permutation-invariant version. The metric dmu is
considered in the literature on mutually unbiased bases [27].
In Publication IV, we show also that constraining the Grassmann man-
ifold to satisfy the EP constraint reduces the dimensionality of the spher-
ical embedding from n2t − 1 to n2
t − nt.
3.3 Codebook Designs
Several criteria have been investigated in the literature to design good
MIMO precoding codebooks. With i.i.d Rayleigh fading, the right eigen-
vectors of the channel are uniformly distributed over the Stiefel manifold
according to the Haar measure [45, 53]. As the set of eigenvectors forms
a representative in a flag manifold, the set of equivalence class is also
Haar distributed over this flag manifold. The main idea is to target uni-
form codebooks over the manifold. For this we need to define a notion
of uniformity. Several mathematically non-equivalent standard criteria
exist. From the point of view of precoding performance metric such as in-
formation rate, they however are almost optimal, and thus roughly equiv-
alent [23, 56, 71, 72, 95]. For uniformity, we will consider both distortion
and discrete arrangement criteria.
Consider a codebook C = {C1, . . . ,Cncb} ⊂ VC
nt,ns. Depending on the dis-
tance dc considered, the codebook is treated as a discretization of the cor-
responding manifold M. Consider a random source V on manifold (M, dc),
16
Flag Manifolds: Code Designs and Geometry
and define the associated quantization map
qc(V) = arg min1≤i≤ncb
dc(V,Ci). (3.14)
In quantization theory, a standard criterion is to minimize the average
distortion of a codebook, i.e. the average squared quantization error,
DM(C) = E[d2c(V,Cqc(V ))
]. (3.15)
Here, we only consider uniformly distributed source, but this criterion is
adaptable for other distributions as well.
In discrete mathematics, the classical approach is to look at the prop-
erties of codebooks as point sets on the respective manifolds. The most
studied criterion is maximizing the minimum distance [19,74]
δM(C) = arg min1≤i,j≤ncb
dc(Cj ,Ci). (3.16)
This problem is often referred as a packing problem and is related to
Tammes problem on the sphere.
A generalization of the packing problem is the Thomson problem where
one has to maximize the p-mean distance [74]:
Mp(C) =⎛⎝ 2
ncb(ncb − 1)
∑1≤j<k≤ncb
dc(Cj ,Ck)p
⎞⎠1/p
. (3.17)
Typical values of interest are p = −1 and −2.
3.4 Lloyd’s Algorithm and Centroids
Lloyd’s algorithm aims to construct a codebook with minimum average
distortion. It comprises two key steps:
Nearest Neighbor rule: Partitioning of the manifold according to the
codebook in ncb Voronoi cells {R1, . . . ,Rncb} defined by
Rk = {V ∈ M| k = qc(V)}. (3.18)
Centroid Computation: Finding the centroids of each Voronoi cell Rk
given by
Zk = arg minV∈M
E[d2c(V,V) | V ∈ Rk
]. (3.19)
The algorithm consists of iterating these two steps where the former code-
book is replaced by the set of computed centroids.
Generally, the centroid of a Voronoi cell is approximated through ex-
haustive search. Meanwhile, due to the treatment of the manifolds as
17
Flag Manifolds: Code Designs and Geometry
subspaces of Euclidean hyperspheres with canonical extrinsic distances,
the centroids may also be computed in closed-form. As the manifolds are
closed continuous surfaces in an Euclidean space, a centroid of a Voronoi
region is obtained from the orthogonal projection onto the manifold of its
center of mass in the ambient space [26]. Accordingly, we provide an exact
centroid computation for the Stiefel manifold in Publication VII, where
an orthogonal projection of any complex matrix to the Stiefel manifold is
given by the polar decomposition following [31,41]. A closed-form solution
was already known for centroids in the Grassmann manifold [58]. While
the closed-form Grassmannian centroid computation is usually proven
differently in the literature, it could also be derived using the argument
in [26]. The centroid corresponds to the closest projection matrix of a
given rank to the center of mass in the ambient space of Hermitian ma-
trices. For deriving centroids in other flag manifolds, the main difficulty
lies in expressing the orthogonal projection. In Publication VIII, the cen-
troid for simple permutation-invariant flag manifolds was approximated
through projection in the embedding space of direct product of Grassman-
nians, a similar approach can be found in [77]. We conjecture that this is
the true centroid.
3.5 Density and Packing Bounds
In the last decade, basic coding-theoretic results estimating the relation-
ship between the cardinality and the minimum distance of codes in Grass-
mann and Stiefel manifolds have been widely studied [7–10,21,22,36,40,
49, 50]. The Hamming bound, a standard coding bound, is related to the
notion of density of codes. In Publications I and II, we discuss the density
of codes in Grassmann and Stiefel manifolds equipped with a chordal dis-
tance. There are two main difficulties in evaluating the density of codes in
these spaces: 1) evaluating the (normalized) volume of a ball 2) estimating
the kissing radius of codes.
1) Volumes: The authors of [22] were able to derive closed-form expres-
sion on the volume of a small ball in Grassmannians under the chordal
distance. A more general approach was used in [49, 50] where a power
series expansion of the (unnormalized) volume of small ball valid for any
Riemann manifold [33] was leveraged. This provides a powerful tool—in
order to obtain a normalized volume expansion, it suffices to divide by the
overall volume of the manifold. The volume of the manifold depends of
18
Flag Manifolds: Code Designs and Geometry
10010−10
10−8
10−6
10−4
10−2
100
chordal distance
volu
me
Unitary group Unt
SimulationTheory
nt=1
nt=2
nt=3 nt=4
nt=5
(a) Unitary groups
100
10−6
10−4
10−2
100
chordal distance
volu
me
Stiefel manifold Vnt,ns
SimulationTheory
2,1
3,1
4,1
4,23,2 4,3
(b) Stiefel manifolds
Figure 3.2. Volume approximation of balls in Stiefel manifolds and unitary groups com-pared to simulation.
the Riemannian metric defining the notion of distance. There is an equiv-
alent intrinsic Riemannian metric corresponding to the extrinsic chordal
distance. In fact, from the Nash embedding theorem, every Riemannian
metric can be seen as being induced by an appropriate Euclidean embed-
ding [65]. For the case of the Stiefel manifold, two non-equivalent Rie-
mannian metrics are often considered [28], one realized from the space of
rectangular unitary matrices and the other from the interpretation of the
manifold as a quotient space of the unitary group embedded in the space of
square unitary matrices. It appears that in the majority of the literature,
the volumes of the manifolds do not correspond to any of these two met-
rics. Indeed, the volume element is unique up to a non-vanishing scaling
factor which is often dismissed, as it can be absorbed in the overall nor-
malization. Here, the volume expansion imposes the scaling of the metric
and volume element. For accurate normalization of the expansion of [33],
the volume of the manifold should be calculated with the same metric. A
discussion and clarification of different conventional normalizations of the
volume of the unitary group is provided in [98]. In [49] the volume of the
complex Stiefel manifold is computed for the geodesic distance induced by
the quotient geometry. In Publication II, we address the problem when
considering the typical chordal distance induced by the rectangular ma-
trix embedding, which leads to an expression of the volume differing from
the ones previously derived in [49] or [40]. The exact volume of the Grass-
mann manifold with chordal distance is also computed. A corresponding
volume computation and mindmap is presented in Fig. 3.3. The exactness
of the volume is illustrated in the estimation of volume of small ball for
the Stiefel manifolds and unitary groups in Fig. 3.2.
19
Flag Manifolds: Code Designs and Geometry
Figure 3.3. Choice of distance and its impact on volume computation and coding boundsfor Stiefel manifolds.
20
Flag Manifolds: Code Designs and Geometry
2) Kissing Radius: As we are considering extrinsic distances, there is
room for improvement of the standard Hamming bound [9, 36] through
the notion of density. We address this problem for the Grassmann mani-
fold in Publication I and for any spherically-embedded manifold in Publi-
cation II. The density of a code is defined as the fraction of the manifold
covered by ‘kissing’ balls of equal radius centered around the codewords.
The kissing radius problem addresses the following question: if two points
are at chordal distance δ, how far is the midpoint from the extremities. For
a geodesic distance, the answer is simply δ/2; with an extrinsic distance,
this is greater than δ/2. While there is a unique and exact answer to this
question on a sphere, this is not the case for flag and Stiefel manifolds
with chordal distances, because in general these spaces are not two-point
homogeneous. For these, the kissing radius cannot be determined solely
from the minimum distance. We provide upper and lower bounds on the
kissing radius as a function of minimum distance for the Grassmann and
Stiefel manifold in Publications I and II.
3) Packing Bounds: Bounds on the kissing radius leads to a refine-
ment of the standard Hamming bound for flag and Stiefel codes. Based
on this, Publication I provides new bounds on the minimum Grassmann
distance. In Publication II, we provide a similar bound for any spherically-
embedded manifold generalizing previously known bounds on codes in the
unitary group [36] to the Stiefel manifold:
Given a code of cardinality ncb and minimum chordal distance δM in a
manifold M isometrically embedded in SD−1(R), we have
δ2M ≤ 4r2n − r4nR2
, (3.20)
where rn is solution of μ(B(rn)) =1
ncband μ(B(rn)) is the normalized vol-
ume of a ball of radius rn.
Applying this result to the unitary group leads to the bound [36, Theo-
rem 2.4]. A tighter bound is provided for a small range of large distances
by [36, Corollary 2.9]. In Publication I, for Grassmann manifolds, we were
able to improve the bound by
δ2g ≤ 4r2n − r4nns
, (3.21)
which is a generalization of the bound in [96] valid only for line packing
(ns = 1).
The Hamming bound is rather loose especially for small codes. For
low cardinality, Rankin bounds [70] related to the spherical embeddings
21
Flag Manifolds: Code Designs and Geometry
are known to be tight when applied to Grassmann codes [18]. In a D-
dimensional sphere, the optimum packings for up to D + 1 points corre-
sponds to simplices. From D + 2 to 2D points, the optimal configurations
are subsets of an orthoplex. They are thus referred as simplex and ortho-
plex bounds, respectively [18].
As discussed in Publication IV, the smaller embedding provided for EP
codebooks leads to a modification of the range of the Rankin bounds. The
corresponding EP orthoplex bound is used for proving the optimality of
some of the packings presented in Publication IV.
22
4. Grassmannian Pakings for 2-TxMIMO
The lowest dimensional flag manifolds (nt = 2) are very specific cases. We
have FC2,2
∼= FC2,1
∼= GC2,1, which further reduces to the real unit sphere
FC2,2
∼= S2. It follows that designing codebooks in FC2,2
∼= GC2,1 is equivalent
to designing spherical codes. In this chapter, we discuss explicit represen-
tation of spherical codes in 2-by-1 vector form for coding in GC2,1; equivalent
2-by-2 matrix codes for FC2,2 can be obtained by pairing an orthogonal com-
plement of each codeword. In addition, a 2× 2 unitary matrix generating
a point in FC2,2 can be seen as two ordered antipodal points, or equiva-
lently an oriented line. It follows that �FC2,2 is the set of spherical antipodal
points, or equivalently the set of lines in 3D, also known as the real Grass-
mannian GR3,1
∼= �FC2,2 [18]. Codebooks in �FC
2,2 can thus be constructed by
leveraging results from known antipodal spherical codes [57].
4.1 Grassmannian Codebooks from Spherical Arrangements
The Grassmann manifold GC2,1 is by definition the complex projective line
CP1. From the Hopf fibration of the unit 3-sphere as a circle bundle over
CP1 [12, Ex.17.23], we have
GC2,1 = CP1 ∼= S3
S1= S2. (4.1)
For the explicit form of the isomorphism we parameterize the unit vector
w, a generator of the equivalence class [w] ∈ GC2,1, as
w(θ, φ) =
⎛⎝ cos θ
eiφ sin θ
⎞⎠ . (4.2)
By setting the range of θ and φ to [0; π2 ] and [0; 2π], respectively, we fully
describe the Grassmannian. Interpreting (θ, φ) directly as spherical coor-
dinates, these would describe a hemisphere. A simple morphism from a
23
Grassmannian Pakings for 2-Tx MIMO
Figure 4.1. The real Grassmannian GR2,1 is the set of lines in R2, or equivalently the set of
antipodal points on the unit circle. Taking only representatives on the upperhemisphere, an isometry to a circle is obtained by doubling the parametricangle θ and reducing the radius to one half.
hemisphere to the whole sphere can be obtained by doubling the angle θ.
The irrelevance of φ for θ = 0 and π2 in (4.2) leads to the isomorphism.
Moreover, with the chordal distance considered, GC2,1 is isometric to the
real sphere (of radius one half). Isometry implies that discretization prob-
lems on (GC2,1, dg) can be addressed on the the real sphere S2. Any spheri-
cal code can be transformed to a Grassmannian codebook by applying the
corresponding simple change of variables. Cartesian coordinates are first
converted to spherical coordinates (ϑ, φ) and the latitude is divided by two
(θ = ϑ2 , φ). A generator of the corresponding Grassmannian line is then
obtained by using (θ, φ) in (4.2). As a result, the chordal distance between
two Grassmannian lines is half the distance between the respective spher-
ical codewords. We illustrate this isometry for φ = 0 which corresponds to
the real Grassmannian GR2,1
∼= S1 isomorphic to a circle in Fig. 4.1.
The problem of distributing a certain number of points uniformly over
the surface of a sphere has been thoroughly studied [74]. Different criteria
on the mutual distances among the codewords have been extremized in
the literature, with motivation often arising from chemistry, biology and
physics [19, 74, 92]. For convenience, solutions are often described as the
vertices of a convex polyhedron.
The Tammes problem is the problem of placing ncb points on a sphere
so as to maximize the minimum distance, also referred to as spherical
packing, is a specific case of spherical arrangements [74]. It follows that
Grassmannian line packing in GC2,1 is equivalent to the Tammes problem.
In Publication III, we have thus construct codebooks and leverage existing
results from the spherical code literature by using the isometry.
24
Grassmannian Pakings for 2-Tx MIMO
0 5 10 15 20 25 30 350.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Number of points
Squ
ared
min
imum
cho
rdal
dis
tanc
e
Squared minimum distance for packing in GC2,1
Simplex boundOrthoplex boundFejes Toth boundHamming−type bound (3.21)Optimal packings (proven)Best known packingsModified Lloyd algorithm [88] Brute−force search [53]
Figure 4.2. Best known squared minimal chordal distance for packings in GC2,1.
The Rankin bounds [70] and the Fejes Tóth bound [30] applies to the
minimum distance of packings in GC2,1. The Fejes Tóth bound is specific
for the 2-sphere. It is tighter than the Hamming-type bound (3.21) which
in this case reduces to the bound in [96]. These bounds lead to proof of
optimality of some packings. Optimum packings of ncb points on a sphere
have been found for ncb ≤ 12 and ncb = 24 [17,30], with optimality proven
geometrically. For ncb up to 130, the best known sphere packings are avail-
able at Sloane’s webpage [84]. Fig. 4.2 shows the achieved minimum dis-
tances of the corresponding Grassmannian packings along with bounds,
compared to numerical results from [94,96] where a modified Lloyd search
algorithm was used and to results from [56] using brute-force computer
search.
The problem of maximizing the generalized p-mean of the mutual dis-
tances among the codewords can be called the generalized Thomson prob-
lem. It is the counterpart of a spherical arrangement problem which, due
to its relevance to physics, is often formulated as the minimization prob-
lem of the Riesz s-energy for s > 0. It is remarked in [74] that on S2 this
problem is only interesting for p < 2.
Some values of p have attracted special interest. The case p = −1 (some-
times also p = −2) is known as the (standard) Thomson problem. Solu-
tions referred to as Fekete points have been found for ncb = 2–4, 6, 12 [25].
Another distinguished problem is the problem of maximizing the product
of the distances, known as Whyte’s problem. This occurs when p → 0 and
can be restated equivalently as minimizing the logarithmic energy. So-
25
Grassmannian Pakings for 2-Tx MIMO
lutions referred to as logarithmic points have been found for ncb = 2–6,
12 [25]. The limiting case p → −∞ is the Tammes Problem discussed
above.
These problems are not in general solved by identical arrangements.
However due to the high symmetry of the optimum solutions of Tammes
problem for 2–4, 6 and 12 points, these cases are conjectured to provide
general solutions [25, 74, 92]. The principal approach to solve these prob-
lems on S2 has been to use extensive numeric computations, especially
in high cardinality. Results may be found at [38, 84] for p = −1 and −∞respectively, and at [13] for p from 0 to −12.
In [37], a library of ncb-point arrangements on a sphere that maximize
the volume of the convex hull is also available. These may also be used as
a basis for constructing precoding codebooks.
4.2 Low Implementation-Complexity Codes
Most solutions of spherical arrangement problems are vertices of polyhe-
dra with a high degree of symmetry which makes the derivation of closed-
form Grassmannian codebooks possible. One benefit of having geometric
insight on the codebooks, and the corresponding analytical handle on their
design, is that suitable rotations can be found by geometric inspection.
Such rotations can be used to simplify the representation of the codebook.
This is beneficial from several perspectives. First, the codebook can be
rotated so that it can be realized with a minimum number of different
complex numbers without impairing performance. Typically, selection of
the precoding codeword is done at the receiver by exhaustive search over
all codewords in the codebook. Codebooks with arbitrary complex entries
result in many complex multiplications at the receiver. Reduced comput-
ing complexity, as well as reduced storage, is possible by constraining the
entries to a finite alphabet. Also, analytic control on the codebooks may
be used to select how the codebooks distribute power across the anten-
nas. Finally, analytic control of the codebooks, together with geometric
intuition, allows investigating non-optimum codebooks, with possibly dif-
ferent symmetry properties than the optimum ones, in order to balance
performance, storage and computing complexity.
In Publication III, we derived simple closed form codebooks from spher-
ical arrangements with up to 5 bits. Of particular interest are the code-
books with ncb = 2, 4, 8 and 16 codewords, i.e. the 1-, 2-, 3- and 4-bit
26
Grassmannian Pakings for 2-Tx MIMO
ncb = 2 Digon⎡⎣ 1√
2±1√
2
⎤⎦
ncb = 4 Tetrahedron[α+
±α−
] [α−
±iα+
]
with α± =√
16(3±√
3)
ncb = 8 Square antiprism[β+
±iβ−
] [β+
±β−
][
β−±γβ+
] [β−
±iγβ+
]
with γ = ei π4
β± =√
12± 1
2
√1+2
√2
ncb = 16[ζ1+
γ2nζ1−
] [ζ1−
γ2n+1ζ1+
][
ζ2+
γ2n+1ζ2−
] [ζ2−
γ2nζ2+
]
with γ = ei π4 , n = 0, 1, 2, 3
ζ1± =
√1±ξ1
2, ξ1 ≈ 0.782
ζ2± =
√1±ξ2
2, ξ2 ≈ 0.236
Figure 4.3. Digon, tetrahedron, square antiprism and 4-bit spherical arrangement.
codebooks. These polyhedra are depicted in Fig. 4.3.
The codebooks in Fig. 4.3 have been rotated in order to decrease search
and storage complexity. To illustrate the implementation benefit of the
closed-form representation, Table 4.1 gives a comparison in terms of the
required number of multiplications and storage bits between random, or
numerically found codebooks, and the 1-, 2-, 3- and 4-bit codebooks of
Fig. 4.3. The required number of complex entries generating the code-
books has been decreased by rotating them so that several points are
on the same latitude. Furthermore, if longitudinal separation between
points on the same latitude are π, π/2, π/4 or a multiple of those, some
27
Grassmannian Pakings for 2-Tx MIMO
Table 4.1. Implementation complexity.
Number of multiplications Storage bits
ncb Proposed CBs Random CBs Proposed CBs Random CBs
2 0 12 2 6Kb
4 4 28 Kb + 21 14Kb
8 6 60 2Kb + 28 30Kb
16 6 124 4Kb + 42 62Kb
complex multiplications can be reduced either to a sign change, a swap be-
tween the real and imaginary parts, additions, or a combination of such.
Additionally, complexity of a codebook can be slightly decreased by scal-
ing it so that the first entry of the first codeword is equal to one. Taking
the Tetrahedron codebook as an example, this gives {(1, ±c), (c,±i )} with
c = α−/α+, thus only one real value, c, needs to be stored, and only four
real multiplications are needed in total. On the other hand, if the en-
tries of the codebook are arbitrary complex numbers, each inner product
between two vectors requires height real multiplications, and storage of
four real values. In summary, with a random codebook of ncb codewords,
the required number of multiplications is 4(2ncb − 1), and the number of
bits required for storage is 2(2ncb − 1)Kb, where Kb is the number of bits
needed to represent a real number.
4.3 Quantization Error Analysis
Since we have designed codebooks analytically, we may be able to compute
the pdf and the average of the squared quantization error in closed form.
For beamforming, the squared quantization error corresponds to a SNR
loss. This can be computed from the interpretation as spherical code The
pdf of the quantization error is given by
fd2g(z) =1
2π
ncb∑k=1
∫Ck(z)∩Rk
dφk. (4.3)
where Rk is the Voronoi cell of the kth codeword, and
Ck(z) = {v ∈ S2(12) : |v − ck|2 = z}.
is the border of a spherical cap of squared radius z centered at codeword
ci.
28
Grassmannian Pakings for 2-Tx MIMO
The integral in the last equality can be calculated by taking into account
the fact that the discontinuities of Ci(z) ∩Ri belong to the borders of Ri.
The borders of Ri are geodesics which may be expressed by a goniometric
equation [67].
We have performed explicit calculations of fd2g for the codebooks of size
1, 2, 3 and 4 bits of the best known Grassmannian packings provided of
Fig. 4.3. These pdfs are drawn in Fig. 4.4(a). For EP codebooks, analytical
expressions for the pdf of the SNR loss, calculated relative to perfect equal
gain beamforming, are given in [63].
Figure 4.4(a) gives a comparison of the average quantization error ob-
tained from different spherical arrangement. The performance of the
maximum volume configurations coincides with the best results found
by Lloyd’s algorithm. The packing solutions perform slightly worse in
general. In term of precoding performance, this corresponds to a loss in
average SNR of the order of 10−3dB – i.e. the difference is insignificant.
0 0.05 0.1 0.15 0.2 0.25 0.3 0.350
5
10
15
20
Square chordal distance (γloss)
Pro
babi
lity
dens
ity
Probability density function of the square chordal distance (γloss)
N=4, TetrahedronN=8, Square AntiprismN=16, Putatively optimal(Tammes Problem)
0 0.1 0.2 0.3 0.4 0.50
5
10
15
20
Square chordal distance (γloss)
Pro
babi
lity
dens
ity N=2, digon
N=4, SquareN=8, OctagonN=16, Hexadecagon
(a) Pdfs
3 3.2 3.4 3.6 3.8 40
0.01
0.02
0.03
0.04
0.05
0.06
0.07
Cardinality of the codebook in bits
Nor
mal
ized
dis
torti
on lo
g 2 D(W
) −(−
1−lo
g2(N
))
Packing (p → −∞)Thomson (p =−1)Whyte/Log. (p → 0)Maximum volumeLloyd algorithmBound
(b) Relative mean squared quantization error
Figure 4.4. Pdfs and mean squared quantization error from constructed codebooks.
29
Grassmannian Pakings for 2-Tx MIMO
30
5. Flag and Stiefel Orbit Codes
Most good codes of Chapter 4 have a natural interpretation as orbits of
a symmetry group [19, 30]. Orbit constructions for unitary codes has not
been widely addressed [66]. Seminal results exist for spherical codes [81,
82]. In [78], unitary codes were constructed from the representation of
fixed-point-free groups, corresponding to an orbit of the identity element.
Along this line, some Stiefel codes were constructed in [34]. For Grass-
mann codes, few works have been addressing group orbit constructions [15,
18,20].
In this chapter, we use group orbits to construct codes in flag and Stiefel
manifolds. Flag orbits are constructed by acting with a unitary repre-
sentation of a finite group. In the construction, the center of the finite
group has no effect, and thus it is sufficient to consider its inner automor-
phism group. Often, orbit construction generates structured codes with
finite input alphabet, which is beneficial for hardware implementation of
MIMO precoding. In this regards, a construction based on the structure
of extraspecial groups, leading to orbits of Clifford groups, is of specific
interest. Other explicit constructions from projective unitary representa-
tions of finite groups are also described. We also give examples of codes on
the Stiefel manifold constructed as orbits of the linear representation of
projective groups, which are thus expansions of the flag codes considered.
5.1 Orbits of Projective Group Representations
Consider a finite unitary group G ⊂ Unt acting on VCnt,ns
and thus on quo-
tient spaces of it. Given a group G and a initial point [Y], Y ∈ VCnt,ns
, in
the manifold of interest, the orbit of [Y] under the action of G is the subset
G[Y] = {[gY] | g ∈ G}. (5.1)
31
Flag and Stiefel Orbit Codes
We shall see that groups with projective representations are of specific
interest for flag orbit construction. The center of the unitary group Unt
is Z(Unt) = {eiθInt | θ ∈ R} ∼= U1. The projective unitary group is the
quotient of the unitary group by its center PUnt = Unt/Z(Unt). An element
in PUnt is an equivalence class of unitary matrices under multiplication by
a constant phase. If a group can be homomorphically mapped to PUnt , it
is said to have a projective unitary representation. Such can be naturally
understood in terms of a linear representation of Un acting on projection
operators by conjugation.
Given a group G having a faithful irreducible representation in Unt , its
inner automorphism group Inn(G) has a representation in PUnt . In Pub-
lications V and VI, we show that flag orbits of the action of G are orbits of
the action of Inn(G): for any [Y ] ∈ F lCnt;s1,...,sr , we have G[Y] = Inn(G)[Y].
It follows that to construct flag orbit codes, we are primarily interested by
groups having a representation in PUnt .
The cardinality of the orbit code depends on the size of the stabilizer sub-
group of the initial point in G. Initial points with trivial stabilizer leads
to orbit codes of the same cardinality as the group. For manifolds without
permutation equivalence, this holds for almost every point in the mani-
fold. When the group has permutation elements, no point in �FCnt,s1,...,sr has
trivial stabilizer. Initial points leading to orbit codes of size less than the
group size have by definition a stabilizer which is a non-trivial subgroup
of G. These are singularities in the manifold and there is only a finite
number of such codes. Such initial points in F lCnt;s1,...,sr are concatena-
tions of r invariant subspaces of dimension {s1, . . . , sr} of some non-trivial
subgroup of G. In this case, appropriate initial points can be found from
eigenspaces of the matrix representation of the group.
In Publication V, we derive some basic properties of Grassmann orbit
codes and describe explicit constructions of group orbits leading to some
optimum packings in 2 and 4 dimensions. Recall that the Grassmann
manifold GCnt,ns
equipped with the chordal distance is isometrically embed-
ded onto a sphere in an Euclidean space of dimension n2t − 1. Any finite
group represented in PUnt acts on the basis of this Euclidean space and
is a subgroup of the orthogonal group SOn2t−1. Except for nt = 2, where
SO3∼= PU2, SOn2
t−1 is larger than PUnt and thus we cannot realize all the
rotations of the Euclidean space with this projective representation. In
2D, we give explicit constructions of group orbits recovering the optimum
packings of Publication III. For higher dimension, we look for groups
32
Flag and Stiefel Orbit Codes
Table 5.1. Orbit codes in GC2,1 of cardinality ncb and minimum squared distance δ2g .
Group Order ncb δ2g Polyhedron
V4∼= Inn(D8) 4 2 1 Digon (optimum)
4 23
Tetrahedron (optimum)
S3 6 2 1 Digon (optimum)
3 34
Triangle (optimum)
6 12
Octahedron (optimum)
D8∼= Inn(D16) 8 4 1
2Square
8 4−√2
7Square antiprism (optimum)
T ∼= A4∼= Inn(2T ) 12 4 2
3Tetrahedron (optimum)
12√
5−1
2√
5Icosahedron (optimum)
O ∼= S4∼= Inn(2O) 24 6 1
2Octahedron (optimum)
8 13
Cube
24 ≈0.1385 Snub cube (optimum)
Table 5.2. Codes from Clifford group in GC4,2 of cardinality ncb and minimum squared
distance δ2g .
ncb δ2g % HB Comments
30 1 66 Orthoplex (optimum). Similar than in [4].
120 0.75 64 Subset of 320-orbit
320 0.44 46
360 0.5 53
390 0.5 54 Union of 30- and 360- orbits. Similar than in [4].
480 0.32 36 Subset of 1440-orbit
710 0.44 54 Union of 30-, 320-, and 360- orbits
1440 0.2 29
2150 0.2 31 Union of 30-, 320-, 360-, and 1440- orbits
that have a relatively simple action on the basis. The Clifford group em-
ployed in quantum information theory permutes and rotates the basis of
the considered space. It is thus natural to consider the Clifford group for
codebook generation. Furthermore, in Publication V we describe several
constructions arising from the Clifford group in 4D, recovering some codes
from [4]. The results are summarized in Table 5.1 and 5.2. In Table 5.2,
the squared minimum distance of the code is evaluated in percentage of
the Hamming-type bound (3.21). The low-cardinality finite groups used
are the Klein Group V4, the symmetric group S3, the dihedral group D8,
the tetrahedral group T and the octahedral group O. Some codes meet the
Rankin bounds. Other justifications of optimality for codes in GC2,1 can be
found in Publication III.
33
Flag and Stiefel Orbit Codes
Table 5.3. Some (ncb, δ2)-flag orbit codes.
�FC2,2
ncb δ2p δ2mu
3 1 1
4 0.66 0.88
6 0.55 0.8
15 0.19 0.35
�FC4,4
ncb δ2p δ2mu
15 2 2
90 1 1
180 0.59 1
360 1.25 1.75
960 0.5 0.75
1440 0.22 0.40
In Publication VI, we give examples of orbit codes in other flag mani-
folds. Their cardinality and squared minimum flag distances are given in
Table 5.3, where δp and δmu are the minimum distance according to (3.11)
and (3.12), respectively.
Designing codes in the lowest dimensional flag manifolds FC2,2
∼= FC2,1
∼=GC2,1
∼= S2 and in �FC2,2
∼= GR3,1 is again equivalent to designing spherical
codes and in �FC2,2
∼= GR3,1 antipodal spherical codes, respectively. Some op-
timal orbit codes in �FC2,2 can be obtained by pairing antipodals of orbit
codes in GC2,1. Examples of optimal orbit codes are simplices of cardinality
3 and 4, orbits of the octahedral group O, forming a octahedron and a cube
on the sphere. The maximum simplicial configuration, i.e. of cardinality
6, forms an icosahedron, an orbit of the tetrahedral group. As expected,
the mutual unbiasedness distances δmu for these codes match the result
of [18], meeting the Rankin bound. A suboptimal packing of size 15 is also
given as orbit of the icosahedral group A5, inner automorphism group of
the binary icosahedral group 2I. The obtained squared mutual unbiased-
ness distance is 0.35, for comparision the putatively optimum code has
δ2mu ≈ 0.38.
We also provide construction in �FC4,4 from orbits of the Clifford Group.
In this space, code elements are 4 × 4 unitary matrices modulo column
permutations and columnwise rotations. From the eigenvectors of the
group elements, we found some initial points with non-trivial stabilizers
of different orders. The resulting codes with cardinality and minimum
distance are presented in Table 5.3. From the table, one can notice than
the two considered distance functions behave quite similarly except for
the code of size 180. The generator of the 15-points codes is the identity
matrix, i.e. the code corresponds to taking the finite group directly as a
code itself. This code is a collection of 3 maximal set of mutually unbiased
34
Flag and Stiefel Orbit Codes
bases [27]. To a (ncb, δ2p)-codes in �FC
4,4, shown in Table 5.3, there exists a
corresponding (24ncb, δ2p)-code in FC
4,4.
5.2 Extraspecial Group Code Constructions
In Publication IV, a construction of Grassmannian packings related to
representation theory is applied to build implementation-friendly codes
when the number of transmit antennas is a power of a prime number.
The codes are multimodal and can be generated from a finite alphabet
consisting of roots of unity. Moreover, their cardinality can be decreased
in order to meet an equal power per antenna constraint.
The construction is based on a group-theoretic framework for packings
in the real Grassmannian, provided in [15] for dimensions that are pow-
ers of 2. This framework is based on the properties of extraspecial 2-
groups [14]. In Publication IV, this construction is generalized to the com-
plex Grassmannian and for any power of a prime p, based on the exten-
sions to extraspecial p-groups in [14]. For dimensions that are powers of 2,
the construction is a specific case of the constructions presented in [3–6]
for non-coherent MIMO and MIMO broadcasting. Whereas in [3–5] the
construction is generalized to construct larger codes, in Publication IV,
we look for smaller codes with good implementation properties. The gen-
eralization to any prime was also done independently in [73].
The main idea is to look for abelian subgroups that are different repre-
sentations of the same group. The subgroups are actually images of an or-
bit under the action of the Clifford group. As representations of an abelian
group are reducible to one-dimensional representations, one can connect
every subgroup with a corresponding decomposition of the representation
space into orthogonal lines, representable by a unitary matrix. The codes
consist of collections of these orthogonal bases and subspaces spanned by
them. As a by-product, the obtained codes are orbits of the Clifford group,
and can also be constructed as such. The extraspecial group machinery
has some advantages over an orbit construction, though. The sizes of
Clifford groups are large and grow factorially with the matrix dimension,
whereas extraspecial groups are rather small and easier to handle. An ho-
momorphism to vector spaces over finite field can also be used to handle
the construction, which generates results on the sizes of the achievable
codes and their minimum distances. Using the orthoplex bound for the
complex Grassmannian, and some of the codes are shown to be optimum
35
Flag and Stiefel Orbit Codes
complex packings, some codes reach the bound with the maximum num-
ber of points.
By using a subset of the codewords, smaller packings satisfying the
equal-power per-antenna constraint can be constructed. Some of the con-
structions are shown to satisfy this constraint in an optimal manner using
a modification of the Conway-Hardin-Sloane spherical embedding of the
Grassmannian for equal-power per-antenna codebooks.
An example of resulting codebook for four transmit antennas with 1-
and 2-stream transmissions is given in Table 5.4. The codebook splits to
three parts. Codebook A corresponds to antenna subset selection. Code-
book {B, C} satisfies the equal-power per-antenna constraint for 2-stream
transmission, while codebook C satisfies the equal power per-antenna con-
straint for both 1-stream and 2-stream transmissions. In Table 5.4, code-
words are not normalized.
Results related to extraspecial code constructions from nt = 2 to nt = 9
(excluding nt = 6 since it is not a power of a prime) are displayed in Ta-
ble 5.5. Optimality of the code is shown by reaching the orthoplex bound,
codes satisfying the equal power per-antenna constraint are commented
as “EP”, optimality with reference to this constraint is shown by reaching
the corresponding orthoplex bound. For nt = 4, the codes can be found in
Table 5.4. For the cases nt = ns, the codes consists of unitary matrices
belonging to a maximal set of mutually unbiased bases (MS-MUB), see
e.g. [27] for details.
Table 5.4. Codebooks from extraspecial-group framework for 4 transmit antennas.
A
⎡⎢⎢⎢⎢⎣
1 0 0 0
0 0 1 0
0 0 0 1
0 1 0 0
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 1 0 0
−i i 0 0
0 0 1 1
0 0 −i i
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 1 0 0
0 0 1 1
i −i 0 0
0 0 −i i
⎤⎥⎥⎥⎥⎦
B
⎡⎢⎢⎢⎢⎣
1 0 0 1
1 0 0 −1
0 −1 1 0
0 1 1 0
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 0 0 1
0 −1 1 0
1 0 0 −1
0 1 1 0
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 0 1 0
0 1 0 1
0 1 0 −1
1 0 −1 0
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 0 1 0
0 1 0 1
0 i 0 −i
−i 0 i 0
⎤⎥⎥⎥⎥⎦
C
⎡⎢⎢⎢⎢⎣
1 1 1 1
1 1 −1 −1
1 −1 1 −1
1 −1 −1 1
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 1 1 1
1 −1 1 −1
−i −i i i
−i i i −i
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 1 1 1
−i i −i i
1 1 −1 −1
−i i i −i
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 1 1 1
−i −i i i
−i i −i i
−1 1 1 −1
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 1 1 1
−i −i i i
−i i −i i
1 −1 −1 1
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 1 1 1
−i i i −i
1 −1 1 −1
i i −i −i
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 1 1 1
1 −1 −1 1
1 −1 1 −1
−1 −1 1 1
⎤⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎣
1 1 1 1
1 −1 1 −1
−i −i i i
i −i −i i
⎤⎥⎥⎥⎥⎦
36
Flag and Stiefel Orbit Codes
Table 5.5. Example of codes from the extraspecial-group framework. Cardinality is ncb
and minimum squared distance δ2c . For nt �= ns, δc corresponds to the min-imum Grassmann distance δg. For nt = ns, δc corresponds to the minimumpermutation-invariant flag distance δp.
nt ns ncb δ2c Comments
2 1 4 12
EP optimum
2 1 6 12
Optimum
2 2 2 1 EP for any sub-rank
2 2 3 1 MS-MUB
3 1 9 23
EP Optimum
3 1 12 23
Optimum
3 3 3 2 EP for any sub-rank
3 3 4 2 MS-MUB
4 1 32 12
EP / columns of {C}
4 1 60 12
Columns of {A, B, C}
4 2 16 1 EP optimum/ EP for sub-rank / 4× 2 matrices in {C}
4 2 24 1 EP optimum / 4× 2 matrices in {B, C}
4 2 30 1 Optimum / 4× 2 matrices in {A, B, C}
4 4 8 2 EP for any sub-rank / 4× 4 matrices in {C}
4 4 15 2 3 MS-MUB / 4× 4 matrices in {A, B, C}
5 1 25 0.8 EP optimum
5 1 30 0.8 Optimum
5 5 5 4 EP for any sub-rank
5 5 6 4 MS-MUB
7 1 49 67
EP optimum
7 1 56 67
Optimum
7 7 7 6 EP for any sub-rank
7 7 8 6 MS-MUB
8 1 512 12
EP
8 1 1080 12
8 2 896 1 EP
8 2 1260 1
8 4 112 2 EP optimum
8 4 126 2 Optimum
8 8 64 4 EP for any sub-rank
8 8 135 4 15 MS-MUB
9 1 243 23
EP
9 1 360 23
9 3 108 2 EP optimum
9 3 120 2 Optimum
9 9 27 6 EP for any sub-rank
9 9 40 6 4 MS-MUB
37
Flag and Stiefel Orbit Codes
5.3 Expansion to Stiefel Codes
Here, we consider Stiefel orbit codes arising from the linear representa-
tion of the projective groups considered in the previous sections. The codes
are expansions of Grassmann orbit codes as direct products of a Grass-
mannian code and a unitary code. Indeed, the obtained codes are more
than just a central extension of the Grassmann code. The codes obtained
are extensions of Grassmannian codes by a finite group of right unitary ro-
tations. Non-trivial stabilizers are only possible if some non-trivial group
elements have eigenvalue 1. Otherwise, the size of the codebook is of the
size of the linear group considered.
In Publication VI, we give examples of Stiefel codes arising from the
Grassmann codes of Publication V. Their cardinality and minimum dis-
tance are summarized in Table 5.6, where Ng and δg stand for the car-
dinality and minimum distance in the Grassmann manifold, whereas Ns
and δs stand for the cardinality and minimum distance in the Stiefel man-
ifold. Their Stiefel squared minimum distances are evaluated in percent-
age of the Hamming-type bound (3.20).
The Stiefel manifold VC2,1 is isomorphic to the 3-sphere, and these two
spaces can be easily mapped to each other. Codes described for VC2,1 are
thus not new and are only interesting as tutorial examples. Some of the
codes are optimal. An orbit of D8 gives an optimal (8, 2)-orthoplex Stiefel
code. An orbit of S3 leads to an example of optimal joint Grassmannian-
Stiefel packings with cardinality 3, meeting the simplex bounds for both
manifolds. The Stiefel orbit of S3 of cardinality 6 and squared minimum
distance 2 is a suborthoplex and is also optimal. Orbits from the binary
tetrahedral group 2T give an optimal Stiefel (24, 1)-packing, vertices of
the 24-cell. This is a well-known polyhedron in 4D with well-understood
symmetry, and known to lead to an optimal packing [84]. Orbits from the
binary octahedral group 2O lead to a codebook of 48 points on the Stiefel
manifold with squared minimum distance 2−√2 ≈ 0.59, a combination of
the 24-cell and its dual which is also a 24-cell. For comparison, the best
known packing of this size has a squared minimum distance of ≈ 0.62 [84].
Using the central extension 2C2 of the Clifford group in 4D, we describe
some codes in VC4,1 and VC
4,2. In VC4,1, the orbit of a (60, 0.5)- Grassmann code
expands to a (480, 2−√2)- Stiefel code. The orbit of a (480, 3
16)- Grassmann
code expands to a (3840, 2− 52√2)- Stiefel code. In VC
4,2, the optimum Grass-
mann orthoplex orbit generates an extension to a (5760, 4−2√2)- Stiefel
38
Flag and Stiefel Orbit Codes
code. The 320-, 360-, and 1440- Grassmann orbit lead to a (15360, 4−2√2)-,
(23040, 1)-, and (46080, 0.40)− Stiefel code, respectively. All these Stiefel
codes are new.
Table 5.6. Some (Ns, δ2s)-Stiefel orbit codes that are expansions of (Ng, δ
2g)-Grassmann
orbit codes.
Dim Group Order Ng δ2g Ns δ2s %HB
2× 1
D8 8
2 1 4 2 62
2 1 8 2 86
4 0.66 8 0.85 37
S3 6
2 1 6 2 75
3 0.75 3 3 84
6 0.5 6 0.59 22
D16 164 0.5 8 0.59 25
8 0.37 16 0.41 26
2T 24 4 0.66 24 1 81
2O 48 6 0.5 48 0.59 73
4× 1 2C2 266!60 0.5 480 0.59 32
480 0.19 3840 0.23 21
4× 2 2C2 266!
30 1 5760 1.17 41
320 0.44 15360 1.17 47
360 0.5 23040 1 43
1440 0.2 46080 0.4 19
39
Flag and Stiefel Orbit Codes
40
6. Joint Grassmann-Stiefel Codes forProduct Codebooks
We consider a product codebook strategy where a single small codebook
is implemented at the receiver to quantize larger MIMO channels, e.g.
aggregate channels of cooperative MIMO base stations or point-to-point
channels with large antenna configuration. This flexible method that
reuses small point-to-point codebooks has many advantages. Only a sin-
gle per-cell codebook needs to be stored for a fixed transmission rank, re-
ducing design problems to smaller spaces which are typically easier to dis-
cretize. Large product codebooks naturally inherit some implementation
properties from the per-cell codebook, e.g. input alphabet and transmit
power per-antenna.
Focusing on the codebook design under this scenario, we propose a novel
joint Grassmann-Stiefel codebook design aiming at good quantization of
Grassmann and Stiefel manifolds with a single codebook, so that prod-
uct codebook quantization becomes competitive with global Grassmann
quantization. We present a vector quantizer to generate Stiefel codebooks
conditioned on a fixed Grassmannian codebook. Some concrete examples
of analytical joint Grassmann-Stiefel packings are also given. We finally
discuss low-complexity codeword selection methods.
6.1 Product Codebook-Based Precoding
We follow the product codebook principle of [16] for feeding back CSI. This
was proposed in order to accommodate to the possible dynamic number of
cooperating BSs and to deal with heterogeneous path loss effects. Details
on the considered CoMP channel model are in Chapter 2. A per-cell code-
book C = {C1, . . . ,Cncb} of (nt × ns)-Stiefel matrices is shared between
the transmitters and the receiver. This codebook is independent of the
number of cooperating BSs, and large-scale path loss effects. The receiver
41
Joint Grassmann-Stiefel Codes for Product Codebooks
A A
A
A
B
B
C B
B
A
C
C
A B C
AAAA A A
B
B
B
B BA
A
A A
C CA
C
C
C
Per-cell CB
Product CBnbs=2
Product CBnbs=3
Figure 6.1. Illustration of product codebook principle for nbs = 1, 2 and 3. Normalizationis not shown.
quantizes Vss with a product codebook Cpr. The product codebook is a
Cartesian product of the per-cell codebook: Cpr = 1√nbs
C ⊗ · · · ⊗ C, i.e. a
codeword in Cpr is a normalized concatenation of nbs single cell codewords.
This is illustrated in Fig. 6.1. Finally, the receiver feeds back the set of
indexes of the codewords of C that form the selected product codeword.
Recall that with ML receiver, when considering the global (nbs×nt)×ns-
system, the performance of the product codebook is related to its interpre-
tation as a discretization of the Grassmann manifold [23,53,56,61]. While
performance of the product codebook is invariant under right unitary ro-
tations of the product codewords, the unitary invariance does not hold
anymore for the per-cell codebook. The space of quantization is then the
Grassmann manifold GCnt,ns
for only one BS, and the Stiefel manifold VCnt,ns
for the other (nbs − 1)-BSs.
In Publication VII, we show that for a Haar distributed right singular
space Vss, the optimum Stiefel quantization of a component of Vss is also
Haar distributed. For the first BS, C thus has to be a uniform Grassman-
nian codebook, whereas for the remaining BSs, C also has to be a uniform
Stiefel codebook. To construct uniform per-cell codebook, one can directly
extend the standard Grassmann codebook criteria to the Stiefel manifold.
In Publication VII, we motivate our choice of Stiefel distance by showing
that the minimum Grassmann chordal distance of the product codebook
Cpr can be lower bounded by a function of the Grassmann and Stiefel chor-
dal distances of the per-cell codebook C:
δ2g(Cpr) ≥ min
{δ2g(C),
δ2g(C) + (nbs − 1)δ2s(C)n2bs
}. (6.1)
The proof is constructive and the bound is thus often tight.
Using a product codebook strategy results in two non-idealities.
i) A single-cell codebook designed to quantize the Grassmann manifold
42
Joint Grassmann-Stiefel Codes for Product Codebooks
1 1.5 2 2.5 3 3.5 43
3.5
4
4.5
5
5.5
6
6.5
7
7.5
8
Number of BS
Spe
ctra
l effi
cien
cy [b
ps/H
z]
Global Grassmann CBProduct CB with proposed designProduct CBs averaged over StiefelProduct CB with good Grassmannquantization but bad Stiefel quantization
(nbsx4)x2
(nbsx2)x1
(a) 1 bit per-Tx, varying MIMO size.
1 1.5 2 2.5 3 3.5 4 4.5 53
3.5
4
4.5
5
5.5
6
6.5
7
Per−cell codebook size in bits
Spe
ctra
l effi
cien
cy [b
ps/H
z]
Global Grassmann CB Product CB with proposed designProduct CBs averaged over StiefelProduct CB with good Grassmannquantization but bad Stiefel quantization
(2x4)x2
(2x2)x1
(b) 2BS, varying per-cell codebook size.
Figure 6.2. Performance comparison of product codebooks and global Grassmann code-books at 10dB SNR.
GCnt,ns
does not necessarily result in a good quantization of the Stiefel man-
ifold VCnt,ns
.
ii) A residual loss would be also expected compared to a global codebook
quantizing the larger Grassmannian GCnbsnt,ns
, corresponding to the signal
eigenspace of the receiver.
To make the performance of product codebook quantization close to op-
timal, we propose a novel joint Grassmann-Stiefel design of the per-cell
codebook C. An example of the achieved performance of such a design is
illustrated in Fig. 6.2(a) and Fig. 6.2(b). Except when the per-cell code-
book size is equal to one bit, the average performance of the proposed
design is close to that of a global Grassmannian codebook (constructed
here via Lloyd’s algorithm). The gain of the proposed design is illus-
trated by comparison with a product codebook based on the same per-
cell Grassmannian codebook but with (putatively) worst choice of Stiefel
representatives. The mean performance of the Grassmann codebook av-
eraged over all possible Stiefel representatives is also shown. The gap
between the best and worst product codebook increase with increasing
number of BSs, while relatively constant with increasing feedback bits.
When the codebook size grows, product codebook performance averaged
over Stiefel representatives is asymptotically reaching the performance
of global codebook.
6.2 Joint Grassmann-Stiefel Codebooks
In order to have good per-cell codebooks that can be used in product code-
books as discussed above, we propose that a codebook is constructed by
43
Joint Grassmann-Stiefel Codes for Product Codebooks
Square 1√3
⎧⎪⎪⎨⎪⎪⎩⎡⎢⎢⎣1
1
1
⎤⎥⎥⎦⎡⎢⎢⎣
1
−1
1
⎤⎥⎥⎦⎡⎢⎢⎣−1
1
1
⎤⎥⎥⎦⎡⎢⎢⎣−1
−1
1
⎤⎥⎥⎦⎫⎪⎪⎬⎪⎪⎭
Tetrahedron 1√3
⎧⎪⎪⎨⎪⎪⎩⎡⎢⎢⎣1
1
1
⎤⎥⎥⎦⎡⎢⎢⎣−1
1
−1
⎤⎥⎥⎦⎡⎢⎢⎣
1
−1
−1
⎤⎥⎥⎦⎡⎢⎢⎣−1
−1
1
⎤⎥⎥⎦⎫⎪⎪⎬⎪⎪⎭
Figure 6.3. Joint Grassmannian-Stiefel codebook design in toy scenario of real codebookfor 3 transmit antennas. On the upper graph the optimum 2-bit Grassman-nian packing in GR
3,1, a set of 4 antipodal points forming a cube. A Grassman-nian codeword may be represented by any of the two points of same color,lying on a line through the origin. On the lower part, two alternatives of 2-bit Stiefel codebooks generating the above Grassmannian codebook: a squareand a tetrahedron.
first designing a Grassmannian codebook according to standard criteria
such as maximizing the minimum distance or minimizing an average dis-
tortion. Then, the representative in each Grassmannian plane in the code-
book is chosen to optimize a metric on the Stiefel manifold. This means
that we select a good Stiefel codebook conditioned on the codebook being
simultaneously a good Grassmannian codebook.
The joint Grassmann-Stiefel codebook design problem is illustrated by
the toy scenario of building a real codebook of four codewords for a trans-
mission from 3 antennas in Fig 6.3. This leads to a rare example where vi-
sualization of the proposed approach is possible. The real Grassmannian
GR3,1 that needs to be discretized is the set of lines through the origin in the
3D Euclidean space. It can be understood as the set of antipodal points
on the real unit sphere. The corresponding Stiefel manifold is the space
of all 3D unit-norm vectors, and can be understood as the full sphere. A
Grassmannian code is then a set of antipodal points, and choosing a rep-
44
Joint Grassmann-Stiefel Codes for Product Codebooks
Table 6.1. 2-bit Square Codebook for 2Tx antenna and a proposed modified version max-imizing the average of the Stiefel distance between the codewords
Square CB 1√2
⎧⎨⎩⎡⎣11
⎤⎦ ⎡⎣ 1
−1
⎤⎦ ⎡⎣1i
⎤⎦ ⎡⎣ 1
−i
⎤⎦⎫⎬⎭Stiefel-improved CB 1√
2
⎧⎨⎩⎡⎣11
⎤⎦ ⎡⎣ 1
−1
⎤⎦ ⎡⎣−1
−i
⎤⎦ ⎡⎣−1
i
⎤⎦⎫⎬⎭Squared Grass. dist.
Squared Stief. dist.
Square CB Stiefel-improved CB
1
1� 2
1 �2
1 �2
1
1� 2
2
1
1
1
2
1
2
3
3
3
2
3
resentative for every Grassmannian codeword means simply choosing one
of the two antipodal points on the sphere. A Stiefel-codebook, in turn, is
a spherical code. The best four-codeword Grassmannian packing is found
by taking the vertices of a cube – the eight vertices of the cube consist
of four pairs of antipodal points, i.e. four Grassmannian lines. From
this cube, there is four possible non-equivalent four-codeword spherical
codes: for example by taking only points in the upper hemisphere we get
a square, or by taking two points in both upper and lower hemispheres
we get a tetrahedron. Those two alternative Stiefel codebooks generating
the same Grassmannian code are given in Fig 4.3. The best Grassmann-
Stiefel codebook is obtained by taking the vertices of the cube that form
a tetrahedron. It turns out that the vertices of the tetrahedron gives ac-
tually the optimal 4-point spherical (Stiefel) codes under several crite-
ria [69]. In this simple example, it is thus possible to have a codebook
that is simultaneously an optimal Grassmannian and Stiefel packing.
This design can be applied to modify existing low-complexity codebooks
from the literature. In Table 6.1, we give a Stiefel-improved version of
the 2-bit Square Codebook discussed in Publications III and IV, which is
related to the Mode 1 codebook of WCDMA [93] and the LTE codebook for
2-transmit antennas [51]. The modified version is obtained by only chang-
ing the sign of the third and last codeword. The squared Stiefel distances
between the codewords of the proposed codebook are either 2 or 3, while
for the original codebook they were either 1 or 2. This codebook has been
45
Joint Grassmann-Stiefel Codes for Product Codebooks
found by brute-force search over QPSK alphabet for the three last code-
words. Furthermore, simulations over all possible phases suggest that
this is putatively the best codebook conditioned on the original codebook
maximizing the p−mean Stiefel distance for p = 1, 2,−1 and −2.
6.3 Lloyd-type Algorithm for Joint Grassmann-Stiefel Codebook
A Lloyd-type algorithm to generate a low-distortion Stiefel codebook con-
ditioned on a Grassmann codebook is presented in Publication VII.
The algorithm is a non-trivial generalization of Lloyd’s algorithm, where
a centroid at a given iteration, used to construct a Voronoi cell, is not nec-
essarily the one updated to the new centroid of the cell. In each Voronoi
cell of the Stiefel manifold, the original codeword is not replaced by the
computed centroid. Instead, the algorithm is looking for the closest code-
word to the the centroid using Grassmann distance rather than Stiefel
distance. As a result, for each Voronoi cell, the updated codeword is not
necessarily belonging to the original Voronoi cell. Then the selected code-
word is replaced by the Stiefel representative in its Grassmann equiva-
lence class closest to the centroid. Indeed, during a single iteration some
codewords can be updated several times and some others not at all. This
occurs in particular in the first iterations. This phenomenon is related to
the non-trivial embedding of the Grassmannian to the Stiefel manifold,
and is crucial for convergence. Simulation results show that the proposed
algorithm converges well.
This feature is illustrated in Fig 6.4 for a toy scenario of choosing the
Stiefel representative of a real Grassmann codebook in GR2,1. The Stiefel
manifold in this case is the unit circle S1, and the Grassmannian is the set
of lines through the origin in 2D, or pairs of antipodal points on a circle.
At Step 1) , the Stiefel representatives of the three Grassmannian lines
have been given in the right half circle. In Steps 2) and 3), the algorithm
generates a random source and partitions the Stiefel manifold based on
a nearest neighbor rule. The Stiefel Voronoi cells corresponding to these
codewords are depicted in blue, red and orange. Non-trivial differences
as compared to the conventional Lloyd’s algorithm can be seen in Step
4) where the algorithm sequentially computes a centroid and updates a
codeword. The centroid of the orange Voronoi cell is depicted in a). It
happens to be closer to the red Grassmannian line than to the orange
one as depicted in b). Thus in c)-d) we update the Stiefel representative
46
Joint Grassmann-Stiefel Codes for Product Codebooks
1) 2)-3)
k = ’orange’
Centroidi = ’red’
Closest Grass. line Update ’red’
R
Converge to
a) b) c)-d)
4)
4)k=’red’
k=’blue’
5)
Figure 6.4. Illustration of the Lloyd-type algorithm for a real Stiefel codebook in VR2,1
∼=S1 conditioned on a real Grassmannian codebook in GR
2,1.
of the red line to this centroid, not the representative of the orange line.
Next, if we consider the centroid of the red Voronoi region, we update the
representative of the orange line, whereas the centroid of the blue region
leads to the representative of the blue line being fixed. As a consequence,
we have found the optimum three-element Grassmann-Stiefel packing in
5).
6.4 Codeword Selections
For a given codebook, optimum quantization of the nbsnt × ns channel
eigenspace requires exhaustive search over the codebook. Here, the prod-
uct codebook has of cardinality nnbscb . This leads to exponential complexity
w.r.t the number of BSs O (nnbscb
). Complexity can be reduced from O (
nnbscb
)to O (nbsncb) by selecting per-cell components rather than selecting jointly
the product codeword [16, 86]. In [16], a lower-complexity selection al-
gorithm trading performance against complexity is proposed, based on
two successive exhaustive searches of size nbsncb and knbs , where k is the
cardinality of preselected per-cell sub-codebooks. Independent and serial
selection are proposed for single-stream beamforming in [42,86].
In Publication VII, we discuss five different codeword selection princi-
ples for multi-stream product codebooks. Two joint selection methods of
47
Joint Grassmann-Stiefel Codes for Product Codebooks
complexity O (nnbscb
)are considered, followed by three selections of com-
plexity O (nbsncb).
• Joint codeword selection is done by selecting the product codeword that
minimizes the global Grassmann distance to Vss [16]. The complexity is
O (nnbscb
).
• Joint codeword selection with transformed codebook improved the pre-
vious selection method in case of path loss imbalance [86] by borrowing
the idea of transformed codebooks for spatially correlated channel [54].
The complexity remains O (nnbscb
).
• Independent Grassmann codeword selection was an alternative proposed
in [86]. Here, each single cell channel component matrix is quantized in-
dependently using the Grassmann chordal distance. This method leads
to a loss of performance as it does not take into account the phase ambi-
guity between the components of the optimum precoding vector as rec-
ognized in [86], or the more general unitary matrix ambiguity. The com-
plexity is reduced to O (nbsncb).
• Independent Grassmann-Stiefel codeword selection: In order to quantize
the per-cell channel components independently and efficiently, the uni-
tary matrix ambiguity between the different channels should be taken
into account. We suggest that first the strongest channel is quantized
using the Grassmannian distance. Then the unitary rotation R not seen
by this Grassmannian codeword selection can be found by performing a
polar decomposition. The channels from the other BSs, with the rota-
tion R taken into account, are then quantized using the Stiefel distance.
Complexity is O (nbsncb).
• Serial Codeword selection: This method borrows the main idea of serial
selection from [42], adapted here to perform codeword selection with a
transformed codebook in a sequential manner. The strongest channel is
first quantized as in the previous independent selection method. Then
the per-cell components are selected sequentially taking into account
the large-scale channel components. Complexity is O (nbsncb).
Figure 6.5 depicts the variation of performance when the large scale
48
Joint Grassmann-Stiefel Codes for Product Codebooks
0 0.2 0.4 0.6 0.8 1
2
2.5
3
3.5
4
4.5
Large scale path loss imbalance α2/α1
Spe
ctra
l effi
cien
cy [b
ps/H
z]
Perfect precodingJoint transformedJointIndep Grass−StiefSerialIndep Grass
(2x4)x2
(2x2)x1
Figure 6.5. Spectral efficiency of 4 × 1 and 8 × 2 systems using 2 × 1 and 4 × 2 code-books respectively, as a function of large scale path gains imbalance. Thestrongest channel is fixed at a SNR of 6 dB. Codebooks with one feedback bitper transmit antenna. The legend indicates the codeword selection methods.Performance of original codebooks is represented in dashed lines, while forStiefel-improved versions, in solid lines.
path gain for the first and the second BS are different. The lower curves
represent 2 cooperative BSs with 2 Tx antennas with 2-bit Square Code-
book and its Stiefel-improved version of Table 6.1. The upper curves rep-
resent two 4-antenna BSs serving a 2 antenna user with 4-bit C-Codebook
from Publication IV and Stiefel-improved version by the proposed Lloyd-
type algorithm. The graph can be interpreted as the performance depend-
ing of the position of the user, from the center of the cell to the cell edge.
The Stiefel-improved codebooks lead to better performance for all code-
word selection methods except independent selection with Grassmann
distance. Using the Stiefel distance consequently improves performance
of independent selection. The serial selection offers slightly better perfor-
mance than independent Grassmannian-Stiefel selection. Both the inde-
pendent Grassmannian-Stiefel selection and the serial selection perform
close to joint selection. The performance gap between joint selection and
independent and serial selection first reduces and then is reversed when
the large scale path gain imbalance grows. With imbalance, joint selec-
tion is not optimal anymore as it quantizes the channel components with
equal weight. Transforming the codebook cures this problem, and gives
the overall best selection method: matching the performance of indepen-
dent selections for large imbalance and joint selection for no imbalance.
49
Joint Grassmann-Stiefel Codes for Product Codebooks
50
7. Flag Codebooks for MIMO Systemswith Linear Receiver
In point-to-point communications with maximum likelihood (ML) receiver,
the performance of a unitary precoding codebook depends on the distance
properties of the Grassmannian planes generated by the codebook. This
has led to the well-known Grassmannian codebook design [55] where the
codebook is understood as a discretization of the Grassmann manifold.
While Grassmann precoding has attracted much attention, other trans-
mission scenarios or constraints may lead to the need to design codebooks
in other flag manifolds.
We illustrate this by focusing on precoding for MIMO systems with
a linear receiver, such as a zero-forcing (ZF) or minimum mean square
(MMSE) receivers. With a linear receiver, the Grassmannian precoding
design used for ML receivers is not anymore appropriate [57]. In Publi-
cation VIII, the spaces of interest are shown to be simple permutation-
invariant flag manifolds.
Of specific interest is the case when the number of streams, and the
number of receive and transmit antennas are the same. In this set up, the
corresponding Grassmannian collapses to a single point making Grass-
mannian precoding irrelevant. Accordingly, precoding does not improve
the information rate with ML receiver. On the other hand, with linear re-
ceiver, gain may be obtained from precoding using the proposed flag code-
book design. Simulations show that this gain is relatively small when only
few feedback bits are used for more than two transmit antennas. This dif-
fers from the behavior of low-rank transmission, where it is known that a
small number of feedback bits can allow near optimal channel adaptation.
51
Flag Codebooks for MIMO Systems with Linear Receiver
7.1 Achievable Information Rates
Let the singular values of H be σ1/21 ≥ . . . ≥ σ
1/2nm with nm = min(nr, nt).
For a fixed precoding vector W ∈ Cnt×ns , the achievable rate depends
on the receiver type. Denote the singular values of the effective channel
Heff = HW by λ1/21 ≥ . . . ≥ λ
1/2ns .
Maximum Rate: Without water-filling and for a given transmission rank
constraint ns, the maximum achievable rate of the system is [23,87]
Ins =
ns∑k=1
log2(1 + γσk) ≤ log2 det(I+ γHHH) = Inm (7.1)
where Inm is the maximum achievable rate without transmission rank
constraint.
ML receiver: With maximum likelihood receiver the achievable rate with
precoding W is
Iml(W) = log2 det(I+ γHeffHHeff) =
ns∑k=1
log2(1 + γλk). (7.2)
We have Iml(W) ≤ Ins . The latter is achievable with Wopt = Vns [23]
where Vns ∈ Cnt×ns is a matrix composed by the right singular vectors of
H corresponding to its ns-largest singular values.
Linear receiver: The receiver employs a linear receiver of the form FHHeff ,
where F = (γHHeffHeff + aIns)
−1. With a = 0, 1, we get a ZF and MMSE
receiver, respectively. The corresponding rate is [39]
Ilr(W) =
ns∑k=1
log2(1 + γk) , (7.3)
where γk = (Fk,k)−1 − a is the post-processing SINR of the k-th data
stream. In general we have Ilr(W) ≤ Iml(W). In [57] it is shown that
there exists a unitary matrix U ∈ Uns such that Ilr(WU) = Iml(W) and
a precoder partitioning is proposed accordingly. As for ML receiver, an
optimum precoding matrix is given by Wopt = Vns .
7.2 Linear Receiver versus ML Receiver
The following difference between linear and ML receivers should be stressed:
1) Full rank n-by-n MIMO: For ns = nt = nr, we have Iml(W) =
Inm for any W ∈ Unt and thus unitary precoding does not change the
transmission rate with ML receiver. With a linear receiver, the rate is a
function of the precoding matrix, and Ilr(W) ≤ Inm .
52
Flag Codebooks for MIMO Systems with Linear Receiver
2) Space of non-equivalent precoder: Let ∼ be the equivalence re-
lation declaring two precoding matrices equivalent, so that W1 ∼ W2, if
and only if I(W1) = I(W2).
ML receiver: The information rate is invariant under any right-unitary
rotations of the precoding codeword: Iml(WU) = Iml(W) for any U ∈Uns . The set of equivalence classes of precoding matrices is exactly the
Grassmann manifold GCnt,ns
.
Linear receiver: The statement above does not hold anymore. The infor-
mation rate with linear receiver (7.3) is invariant under permutations and
phase multiplications of columns of the precoding matrix, Ilr(WDP) =
Ilr(W), for any permutation matrix P ∈ Uns and any diagonal matrix
D ∈ Uns . The set of equivalence classes of precoding matrices is thus the
permutation-invariant flag manifold �FCnt,ns
.
7.3 Codebook Designs
Now we assume that the channel matrix is a random variable and that
the transmitter picks the precoding matrix from a codebook C following a
quantization rule. Given the instantaneous transmission rate I(H,W) of
a precoding matrix W, the average information rate is
I = EH
[I(H,Cq(H))
]. (7.4)
The codebook should be designed as a discretization of the space VCnt,ns
/∼of equivalence classes of precoding matrices. The task thus becomes:
ML receiver: discretize the Grassmann manifold GCnt,ns
,
Linear receiver: discretize the flag manifold �FCnt,ns
.
An optimum quantization q∗(H) = argmax1≤i≤nbI(Ci) is untractable.
For Grassmannian precoding, quantization with Grassmann chordal dis-
tance, qg(Vns) = argmin1≤i≤nbdg(Vns ,Ci), has been considered instead
and shown to be asymptotically optimum [23]. Corresponding good code-
books are thus designed by minimizing the average squared distortion
E[d2g(Vns ,Cqg(Vns ))] [22, 23]. We extend this principle to codebook design
on the flag manifold by replacing the chordal distance dg by the distance
dp defined in (3.11). A corresponding Lloyd’s algorithm providing low-
distortion codebooks is given in Publication VIII.
53
Flag Codebooks for MIMO Systems with Linear Receiver
7.4 Simulations
We illustrate by simulations the performance of the proposed design for
i.i.d Rayleigh channels, using codebooks generated by the Lloyd’s algo-
rithm described in Publication VIII and the flag orbit codes of Chapter 5.
While the codebooks has been designed using the distance metric dp, we
use the optimal quantization metric, i.e. the information rate, for selec-
tion of the codeword at the receiver. Indeed, unlike for the Grassmann
precoding where optimal selection and chordal distance selection show
similar performance, codeword selection on flag manifold with dp injures
a significant performance loss.
Fig. 7.1 shows the information rate for 4 × 2 MIMO systems with ZF
receiver, rank-2 transmission and 4-bit codebooks in �FC4,2 and GC
4,2. The
flag manifold codebook outperforms the Grassmannian with more than
1 dB. The performance of the precoding partitioning scheme of [57] is
also depicted. Its performance falls between the Grassmannian and flag
codebooks. In the partitioning, bits are split equally between Grassman-
nian precoding and 2-stream orthogonalization, and codebooks for both
are generated by Lloyd’s algorithms.
Fig. 7.2 shows the information rate for a full-rank system with an equal
number of transmit and receive antennas. Additionally to codebooks from
Lloyd’s algorithm, simulations are performed with flag orbit codes of Chap-
ter 5. The size of the codes is expressed in bits. In this scenario, as pointed
above, precoding is irrelevant if ML receiver is used, while for a linear re-
10 10.5 11 11.5 12 12.5 13 13.5 14 14.5 154
5
6
7
8
9
10
11
SNR(dB)
Spe
ctra
l Effi
cien
cy (b
ps/H
z)
Perfect PrecodingF(4,2) with 4 feedback bitsG(4,2)+F(2,2), 4 feedback bits (2 each)G(4,2) with 4 feedback bitsNo Precoding
Figure 7.1. Information rate with ZF receiver for nt = 4, nr = ns = 2, four-bit (nb = 16)codebooks designed in flag �FC
4,2 and Grassmann GC4,2 manifolds, and precoding
partitioning [57] with 2 bits for Grassmannian and orthogonalization each.
54
Flag Codebooks for MIMO Systems with Linear Receiver
ceiver, precoding has an effect. For 2× 2 systems, it is possible to recover
most of the gap between no precoding and perfect precoding with a few
feedback bits. However, when the number of antennas increases, e.g. for
nt = nr = 4 in Fig. 7.2, the marginal gain for a small number of bits is
small. In the SNR range depicted, the performance gap between no and
perfect precoding is some 5dB. Precoding with 10 bits recovers only half
of this gap. The codebooks from Lloyd’s algorithm perform slightly better
than the orbit code.
The curves on Fig. 2 of Publication VIII correspond to performance with
codeword selection using the flag chordal distance. Thus, in Publica-
tion VIII the performance is lower compared to Fig. 7.2 where optimum
selection is performed.
10 11 12 13 14 154
4.5
5
5.5
6
6.5
7
7.5
8
SNR [dB]
Spe
ctra
l effi
cien
cy [b
ps/H
z]
perfect precodingOrbit codesLloyd’s CBsno precoding
2 and ≈ 4 bits
1 to 5 bits
(a) Full rank transmission of 2× 2. Codebooks size up to 5 bits.
10 11 12 13 14 156
7
8
9
10
11
12
13
14
SNR [dB]
Spe
ctra
l effi
cien
cy [b
ps/H
z]
perfect precodingorbit codesLloyd’s CBsno precoding
≈ 4/ 6.5/ 7.5/ 8.5/ 10/ 10.5 bits
1 to 6 bits
(b) Full rank transmission of 4×4. Codebooks size up to 10 bits.
Figure 7.2. Information rate of full rank transmission nt = nr = ns with ZF receiver.
55
Flag Codebooks for MIMO Systems with Linear Receiver
56
8. Conclusions
Motivated by applications to codebook-based unitary precoding for limited-
feedback MIMO systems, we have considered coding in flag manifolds
equipped with chordal distances. Codebook-based unitary precoding is
widely employed, and several codebook designs in the literature are exam-
ples of flag manifold discretization. Analytical constructions are consid-
ered to acquire analytical control of the codebooks, so that e.g. providing
low implementation-complexity for practical systems.
We described spherical embeddings of flag manifolds with the corre-
sponding chordal distances. Flag codes are a subclass of spherical codes,
and coding and geometrical problems on flag manifolds are mathematical
problem of independent interest. We have discussed centroid computa-
tion, volume of metric balls, and Hamming-type bounds on Grassmann
and Stiefel codes with chordal distance. Geometry of manifolds depends
on the choice of the Riemannian metric, implying an Euclidean emded-
ding. With a chordal distance, understanding the corresponding embed-
ding enables leveraging results from Euclidean geometry literature.
Code constructions are based on geometric intuition from the lowest di-
mensional flag manifolds, the circle and the real sphere. We have dis-
cussed the problem of designing closed-form codebooks for two transmit
antennas. The problem reduces to a quantization problem on a real 2-
sphere. Utilizing a simple isomorphism, we were able to derive simple
closed form codebooks from spherical arrangements with inherent low
implementation complexity. The discussed code construction is specific
for two transmit antenna systems and cannot be straight forwardly gen-
eralized to higher dimensions.
For more transmit antennas, i.e. higher dimensional codebooks, we con-
sider three different constructions: orbit codes, extraspecial group con-
structions, and product codebooks.
57
Conclusions
We discussed flag orbit codes arising from projective unitary group rep-
resentations. We described few examples in 2D and 4D for the Grassmann
manifold and unitary matrices modulo column permutations and column-
wise rotations. We also described Stiefel orbit codes, as group expansions
of constructed Grassmann orbit codes. For dimensions equal to a power
of a prime, we have presented a construction of Grassmannian packings
related to representation theory of extraspecial groups. We prove the opti-
mality of several of the constructed codebooks with reference to an equal-
power per-antenna constraint. For MIMO precoding, orbit constructions
enable building well structured codebooks. However, an orbit construc-
tion has the drawback of giving very little freedom in controlling the size
of the codes, which often is not an exact number of bits. Also, for many
antenna systems, this requires manipulating very large groups.
We have considered product codebook quantization where codewords
from a single small point-to-point codebook are concatenated to quantize
larger MIMO channels, e.g. channels from cooperative BSs. We have
proposed a joint Grassmann-Stiefel codebook design to diminish the per-
formance gap between product codebook quantization and global Grass-
mannian quantization. We have investigated methods to construct good
Stiefel codebooks conditioned on Grassmannian codebooks. A Lloyd-type
algorithm on Stiefel manifold conditioned on a given Grassmannian code-
book is proposed, as well as some closed-form examples of joint Grassmann-
Stiefel codebooks. For large MIMO system, product codebook framework
is a promising method as it offers very good performance, and also reduces
the codebook design to smaller spaces which are easier to handle.
We finally concentrated on the design of unitary codebooks for MIMO
systems with linear receivers. The correct spaces of quantization are cer-
tain permutation-invariant flag manifolds. The pertinence of the design
principle was illustrated by simulations. In full-rank MIMO with linear
receiver, the capacity is not achieved by default unlike with ML receiver.
Moreover, for more than two transmit antennas, precoding with limited
feedback offers only small gain. Significant amounts of feedback bits are
needed to improve the rates of high-rank MIMO transmission with linear
receivers. This behavior differs significantly from low-rank unitary pre-
coding whose success arises from good marginal gains for few feedback
bits.
Possible directions for future research include extending the volume ball
estimation, kissing radius and Hamming bound to other flag manifolds.
58
Conclusions
Using this volume estimation, it would be possible to generalize bounds
on rate-distortion tradeoff in Grassmann manifolds [22] to Stiefel mani-
folds [49] and other flag manifolds. Then, distortion bounds may be used
for evaluating the information rate of flag precoding. Accordingly, it would
be interesting to quantify analytically the observation of Chapter 7 on
the small marginal gain of precoding for full-rank MIMO with linear re-
ceiver. Alternatively, differential precoding could be investigated to cure
this problem. Finally, another direction of future work is to build new
orbit codes by investigating other finite groups.
59
Conclusions
60
Bibliography
[1] 3GPP TSG . RAN WG1 Meeting #55bis R1-090388, Ljubljana, Slovenia,12-16 January, 2009.
[2] V. Aggarwal, A. Ashikhmin, and A. Calderbank. A Grassmannian packingbased on the nordstrom-robinson code. In Proc. IEEE Inf. Theory Workshop,pages 1–5, Oct. 2006.
[3] A. Ashikhmin and A. Calderbank. Space-time Reed-Muller codes for non-coherent MIMO transmission. In Proc. IEEE Int. Symp. Inf. Theory, pages1952–1956, Sept. 2005.
[4] A. Ashikhmin and A. Calderbank. Grassmannian packings from operatorreed–muller codes. IEEE Trans. Inf. Theory, 56(11):5689–5714, Nov. 2010.
[5] A. Ashikhmin, A. Calderbank, and W. Kewlin. Multidimensional secondorder Reed-Muller codes as Grassmannian packings. In Proc. IEEE Int.Symp. Inf. Theory, pages 1001–1005, July 2006.
[6] A. Ashikhmin and R. Gopalan. Grassmannian packings for efficient quan-tization in MIMO broadcast systems. In Proc. IEEE Int. Symp. Inf. Theory,pages 1811–1815, June 2007.
[7] C. Bachoc. Linear programming bounds for codes in Grassmannian spaces.IEEE Trans. Inf. Theory, 52(5):2111–2125, May 2006.
[8] C. Bachoc, Y. Ben-Haim, and S. Litsyn. Bounds for codes in productsof spaces, Grassmann, and Stiefel manifolds. IEEE Trans. Inf. Theory,54(3):1024–1035, Mar. 2008.
[9] A. Barg and D. Nogin. Bounds on packings of spheres in the Grassmannmanifold. IEEE Trans. Inf. Theory, 48(9):2450 – 2454, Sep. 2002.
[10] A. Barg and D. Nogin. A bound on Grassmannian codes. In Proc. IEEE Int.Symp. Inf. Theory, pages 997–1000, July 2006.
[11] A. Borel. Groupes algébriques. Séminaire N. Bourbaki, 3(121):1–10, 1954-1956.
[12] R. Bott and L. Tu. Differential forms in algebraic topology. Springer –Verlag, 1982.
[13] M. Bowick, C. Cecka, L. Giomi, A. Middleton, and K. Ziel-nicki. Thomson Problem @S.U. published electronically athttp://physics.syr.edu/condensedmatter/thomson/thomson.htm.
61
Bibliography
[14] A. Calderbank, P. Cameron, W. Kantor, and J. Seidel. Z4-Kerdock codes,orthogonal spreads, and extremal Euclidean line-sets. Proc. London Math.Soc., 75(2):436–480, 1997.
[15] A. Calderbank, R. Hardin, E. Rains, P. Shor, and N. Sloane. A group-theoretic framework for the construction of packings in Grassmannianspaces. J. Algebraic Combin., 9:129–140, 1999.
[16] Y. Cheng, V. Lau, and Y. Long. A scalable limited feedback design for net-work MIMO using per-cell product codebook. IEEE Trans. Wireless Com-mun., 9(10):3093–3099, Oct. 2010.
[17] B. W. Clare and D. Kepert. The closest packing of equal circles on a sphere.P. Roy. Soc. Lond. A Mat., 405(1829):329–344, 1986.
[18] J. Conway, R. Hardin, and N. Sloane. Packing lines, planes, etc.: Packingsin Grassmannian space. Exp. Math, 5:139–159, 1996.
[19] J. H. Conway and N. Sloane. Sphere Packings, Lattices and Groups.Springer-Verlag, New York, 2nd edition, 1993.
[20] J. Creignou. Constructions of Grassmannian simplices. eprintarXiv:cs/0703036, Mar. 2007.
[21] J. Creignou and H. Diet. Linear programming bounds for unitary space timecodes. In Proc. IEEE Int. Symp. Inf. Theory, pages 1073–1077, July 2008.
[22] W. Dai, Y. Liu, and B. Rider. Quantization bounds on Grassmann mani-folds and applications to MIMO communications. IEEE Trans. Inf. Theory,54(3):1108–1123, Mar. 2008.
[23] W. Dai, Y. Liu, B. Rider, and V. Lau. On the information rate of MIMO sys-tems with finite rate channel state feedback using beamforming and poweron/off strategy. IEEE Trans. Inf., 55(11):5032–5047, Nov. 2009.
[24] I. Dhillon, R. Heath, T. Strohmer, and J. Tropp. Constructing packings inGrassmannian manifolds via alternating projection. Exp. Math., 17(1):9–35,2008.
[25] P. Dragnev, D. Legg, and D. Townsend. Discrete logarithmic energy on thesphere. Pacific J. Math., 207(2):345–358, 2002.
[26] Q. Du, M. Gunzburger, and L. Ju. Constrained centroidal voronoi tessella-tions for surfaces. SIAM J. Sci. Comput., 24(5):1488–1506, 2003.
[27] T. Durt, B.-G. Englert, I. Bengtsson, and K. Zyczkowski. On mutually unbi-ased bases. Int. J. Quanum Inform., 8(04):535–640, 2010.
[28] A. Edelman, T. Arias, and S. Smith. The geometry of algorithms with or-thogonality constraints. SIAM J. Matrix Anal. A., 20(2):303–353, 1998.
[29] C. Ehresmann. Sur la topologie de certains espaces homogenes. Ann. Math.,35(2):396–443, Apr. 1934.
[30] T. Ericson and V. Zinoviev. Codes on Euclidean Spheres. North-HollandMathematical Library, 2001.
[31] K. Fan and A. Hoffman. Some metric inequalities in the space of matrices.Proc. Amer. Math. Soc., 6(1):111–116, Feb. 1955.
62
Bibliography
[32] G. Foschini and M. Gans. On limits of wireless communication in a fad-ing environment when using multiple antennas. Wireless Pers. Commun.,6(3):311–335, Mar. 1998.
[33] R. Gray and D. Neuhoff. Quantization. IEEE Trans. Inf. Theory, 44(6):2325–2383, Oct. 1998.
[34] G. Han and J. Rosenthal. Geometrical and numerical design of struc-tured unitary space amp-time constellations. IEEE Trans. Inf. Theory,52(8):3722–3735, 2006.
[35] G. Han and J. Rosenthal. Geometrical and numerical design of structuredunitary space-time constellations. IEEE Trans. Inf. Theory, 52(8):3722–3735, Aug. 2006.
[36] G. Han and J. Rosenthal. Unitary space-time constellation analysis: Anupper bound for the diversity. IEEE Trans. Inf. Theory, 52(10):4713 –4721,Oct. 2006.
[37] R. Hardin, N. Sloane, and W. Smith. Maximal volume spherical codes. pub-lished electronically at http://www2.research.att.com/∼njas/maxvolumes/.
[38] R. Hardin, N. Sloane, and W. Smith. Minimal energy ar-rangements of points on a sphere. published electronically athttp://www2.research.att.com/∼njas/electrons/.
[39] R. Heath, S. Sandhu, and A. Paulraj. Antenna selection for spatial multi-plexing systems with linear receivers. IEEE Commun. Lett, 5(4):142 –144,Apr. 2001.
[40] O. Henkel. Sphere packing bounds in the Grassmann and Stiefel manifolds.IEEE Trans. Inf. Theory, 51:3445, Oct. 2005.
[41] N. Higham. Matrix nearness problems and applications. In Applications ofMatrix Theory, Oxford University Press, pages 1–27, 1989.
[42] X. Hou and C. Yang. Codebook design and selection for multi-cell cooper-ative transmission limited feedback systems. In Proc. IEEE Veh. Technol.Conf., pages 1–5, May 2011.
[43] K. Huang, J. Andrews, and R. Heath. Performance of orthogonal beamform-ing for SDMA with limited feedback. IEEE Trans. Veh. Technol., 58(1):152–164, Jan. 2009.
[44] T. Inoue and R. Heath. Kerdock codes for limited feedback precoded MIMOsystems. Trans. Signal Process., 57(9):3711–3716, Sep. 2009.
[45] A. T. James. Distributions of matrix variates and latent roots derived fromnormal samples. Ann. Math. Statist., 35:475–501, June 1964.
[46] M. Karakayali, G. Foschini, and R. Valenzuela. Network coordination forspectrally efficient communications in cellular systems. IEEE Wireless.Commun., 13(4):56–61, Aug. 2006.
[47] I. Kim, S. Park, D. Love, and S. Kim. Improved multiuser MIMO unitaryprecoding using partial channel state information and insights from theRiemannian manifold. IEEE Trans. Wireless Commun., 8(8):4014–4023,Aug. 2009.
63
Bibliography
[48] J. Kim, W. Zirwas, and M. Haardt. Efficient feedback via subspace-basedchannel quantization for distributed cooperative antenna systems with tem-porally correlated channels. EURASIP J. Adv. Sig. Pr., (847296):1–13, 2008.
[49] R. Krishnamachari and M. Varanasi. Volume of geodesic balls in the com-plex Stiefel manifold. In Proc. of Allerton Conf. on Comm. Cont. and Comp.,pages 902–909, Sep. 2008.
[50] R. Krishnamachari and M. Varanasi. Volume of geodesic balls in the realStiefel manifold. In Proc. 42nd Annu. Conf. Inf. Sci. Syst., pages 402–406,Mar. 2008.
[51] P. Lee, J.-K. Han, and J. Zhang. MIMO technologies in 3GPP LTE and LTE-Advanced. EURASIP J. Wirel. Commun. Netw., (302092):1–10, July 2009.
[52] D. Love and R. Heath. Equal gain transmission in multiple-input multiple-output wireless systems. IEEE Trans. on Commun., 51(7):1102–1110, 2003.
[53] D. Love and R. Heath. Limited feedback unitary precoding for spatial mul-tiplexing systems. IEEE Trans. Inf., 51(8):2967–2976, Aug. 2005.
[54] D. Love and R. Heath. Limited feedback diversity techniques for correlatedchannels. IEEE Trans. Veh. Technol., 55(2):718–722, Mar. 2006.
[55] D. Love, R. Heath, V. Lau, D. Gesbert, B. Rao, and M. Andrews. An overviewof limited feedback in wireless communication systems. IEEE J. Sel. AreasCommun., 26(8):1341 –1365, Oct. 2008.
[56] D. Love, H. R.W., and T. Strohmer. Grassmannian beamforming formultiple-input multiple-output wireless systems. IEEE Trans. Inf. Theory,49(10):2735–2747, Oct. 2003.
[57] H. Määttänen, K. Schober, O. Tirkkonen, and R. Wichman. Precoder par-titioning in closed-loop MIMO systems. IEEE Trans. Wireless Commun.,8(8):3910–3914, Aug. 2009.
[58] B. Mondal, S. Dutta, and R. Heath. Quantization on the Grassmann mani-fold. IEEE Trans. Signal Process., 55(8):4208–4216, Aug. 2007.
[59] B. Mondal and R. Heath. Performance analysis of quantized beamformingMIMO systems. IEEE Trans. Signal Process., 54(12):4753–4766, Dec. 2006.
[60] B. Mondal, T. Thomas, and M. Harrison. Rank-independent codebook designfrom a quaternary alphabet. In Proc. Asilomar Conf. on Signals, Systemsand Computers, pages 297–301, Nov. 2007.
[61] K. K. Mukkavilli, A. Sabharwal, E. Erkip, and B. Aazhang. On beamformingwith finite rate feedback in multiple-antenna systems. IEEE Trans. Inf.Theory, 49(10):2562–2579, Oct. 2003.
[62] S. Murray and C. Sämann. Quantization of flag manifolds and their super-symmetric extensions. Adv. Theor. Math. Phys., 12(3):641–710, 2008.
[63] C. Murthy and B. D. Rao. Quantization methods for equal gain transmissionwith finite rate feedback. IEEE Trans. Sig. Proc., 55(1):233 –245, Jan. 2007.
[64] A. Narula, M. Lopez, M. Trott, and G. Wornell. Efficient use of side infor-mation in multiple-antenna data transmission over fading channels. IEEEJ. Sel. Areas Commun., 16(8):1423–1436, Oct. 1998.
64
Bibliography
[65] J. Nash. C1 isometric imbeddings. Ann. Math. (2), 60(3):383–396, Nov. 1954.
[66] F. Oggier. A survey of algebraic unitary codes. In Coding and Cryptology,pages 171–187. Springer, 2009.
[67] J. Oprea. Differential Geometry and Its Applications. Prentice Hall, 2003.
[68] R. Picken. The Duistermaat-Heckman integration formula on flag mani-folds. J. of Math. Phys., 31(3):616–638, Mar. 1990.
[69] R.-A. Pitaval, H.-L. Määttänen, K. Schober, O. Tirkkonen, and R. Wichman.Beamforming codebooks for two transmit antenna systems based on opti-mum grassmannian packings. IEEE Trans. Inf. Theory, 57(10):6591–6602,Oct. 2011.
[70] R. Rankin. The closest packing of spherical caps in n dimensions. Proc.Glasgow Math. Assoc., 2:139–144, 1955.
[71] J. Roh and B. Rao. Design and analysis of MIMO spatial multiplexing sys-tems with quantized feedback. IEEE Trans. Sig. Proc., 54(8):2874–2886,Aug. 2006.
[72] J. Roh and B. Rao. Transmit beamforming in multiple-antenna systemswith finite rate feedback: a VQ-based approach. IEEE Trans. Inf. Theory,52(3):1101–1112, Mar. 2006.
[73] A. Roy. Bounds for codes and designs in complex subspaces. J. AlgebraicCombin., 31(1):1–32, 2010.
[74] E. Saff and A. Kuijlaars. Distributing many points on a sphere. Math.Intell., 19(1):5–11, 1997.
[75] J. Saltz. Digital transmission over cross-coupled linear channels. AT & TTechnical Journal, 64(6):1147–1159, July-Aug. 1985.
[76] K. Schober, P. Janis, and R. Wichman. Geodesical codebook design for pre-coded mimo systems. IEEE Commun. Lett., 13(10):773–775, Oct. 2009.
[77] E. Sengul, H. Park, and E. Ayanoglu. Bit-interleaved coded multiple beam-forming with imperfect CSIT. IEEE Trans. Commun., 57(5):1505–1513,May 2009.
[78] A. Shokrollahi, B. Hassibi, B. Hochwald, and W. Sweldens. Representationtheory for high-rate multiple-antenna code design. IEEE Trans. Inf. Theory,47(6):2335–2367, 2001.
[79] P. Shor and N. J. Sloane. A family of optimal packings in Grassmannianmanifolds. J. Algebraic Combin., 7:157–163, 1998.
[80] T. Shuang, T. Koivisto, H.-L. Maattanen, K. Pietikainen, T. Roman, andM. Enescu. Design and evaluation of LTE-Advanced double codebook. InProc. IEEE Veh. Technol. Conf., pages 1–5, May 2011.
[81] V. Sidelnikov. On a finite group of matrices generating orbit codes on Eu-clidean sphere. In Proc. IEEE Int. Symp. Inf. Theory, page 436, June 1997.
[82] D. Slepian. Group codes for the Gaussian channel. Bell Syst. Techn. J.,47(4):575–602, Apr. 1968.
65
Bibliography
[83] N. Sloane. Neil J. A. Sloane: Home page. published electronically athttp://www.research.att.com/∼njas/index.html#TABLES.
[84] N. Sloane, with the collaboration of R.H. Hardin, W. Smith,et al. Tables of spherical codes. published electronically atwww.research.att.com/∼njas/packings/.
[85] T. Strohmer and R. Heath. Grassmannian frames with applications to cod-ing and communication. Appl. Comput. Harmon. A., 14(3):257–275, 2003.
[86] D. Su, X. Hou, and C. Yang. Quantization based on per-cell codebook incooperative multi-cell systems. In Proc. IEEE Wireless Commun. & Net.Conf., pages 1753 –1758, Mar. 2011.
[87] E. Telatar. Capacity of multi-antenna gaussian channels. Eur. T. Telecom-mun., 10:585–595, 1999.
[88] L. Thiele, M. Schellmann, T. Wirth, and V. Jungnickel. Cooperative multi-user MIMO based on limited feedback in downlink OFDM systems. In Proc.Asilomar Conf. on Signals, Systems and Computers, pages 2063–2067, Oct.2008.
[89] E. Visotsky and U. Madhow. Space-time transmit precoding with imperfectfeedback. IEEE Trans. Inf. Theory, 47(6):2632–2639, Sep. 2001.
[90] S. Wagner, S. Sesia, and D. Slock. Unitary beamforming under constantmodulus constraint in MIMO broadcast channels. In Proc. IEEE WorkshopSig. Process. Adv. Wireless Commun., pages 151 –155, June 2009.
[91] S. Wagner, S. Sesia, and D. Slock. On unitary beamforming for MIMO broad-cast channels. In Proc. IEEE Int. Conf. Commun., pages 1–5, May 2010.
[92] L. Whyte. Unique arrangements of points on a sphere. Amer. Math.Monthly, 59:606–611, 1952.
[93] R. Wichman and A. Hottinen. Transmit diversity in the WCDMA system.Int. J. Wireless Inf. Networks, 6:171–180, 1999.
[94] P. Xia. Multi-Input Multi-Output Communications with Partial ChannelState Information. PhD thesis, University of Minnesota, 2005.
[95] P. Xia and G. Giannakis. Design and analysis of transmit-beamformingbased on limited-rate feedback. IEEE Transactions on Signal Processing,54(5):1853–1863, May 2006.
[96] P. Xia, S. Zhou, and G. Giannakis. Achieving the Welch bound with differ-ence sets. IEEE Trans. Inf. Theory, 51(5):1900–1907, May 2005.
[97] L. Zheng and D. Tse. Communication on the Grassmann manifold: a geo-metric approach to the noncoherent multiple-antenna channel. IEEE Trans.Inf. Theory, 48(2):359–383, Feb. 2002.
[98] K. Zyczkowski and H.-J. Sommers. Hilbert-Schmidt volume of the set ofmixed quantum states. J. Phys. A: Math. Gen., 36(39):10115–10130, 2003.
66
Errata
Publication I
p. 2299, first paragraph, the principal angles are given by the SVD of Y †Z
not Y Z†.
Publication III
The symbol of the border of the spherical cap Ci(z) did not print correctly,
as a result both the cap and its border are defined by Ci(z):
• p. 6597, the last equation should be:
Ci(z) = {y ∈ S2(12) : |y − xi|2 = z}.
• p. 6598, the derivation of equation (24) should read:
A(Ci(z) ∩ Vi) =
∫∫Ci(z)∩Vi
1
2dz dφi
=
∫ z
0
(1
2
∫Ci(z)∩Vi
dφi
)dz.
The pdf of the squared distance fd2c is then obtained by straightforward
differentiation:
d
dz(A(Ci(z) ∩ Vi)) =
1
2
∫Ci(z)∩Vi
dφi,
and finally we have
fd2c (z) =1
2π
N∑i=1
∫Ci(z)∩Vi
dφi.
67
Errata
The integral in the last equality can be calculated by taking into account
the fact that the discontinuities of Ci(z) ∩ Vi belong to the borders of Vi.
Publication IV
• p. 5 Table III, the entries of some codewords are incorrect. A corrected
version is given in Table 5.4.
• p. 3 Proposition 2, and p. 4 Proposition 4, totally singular subspaces
should be understood as totally isotropic subspaces. Singular and isotropic
subspaces was the terminology used in [14] when considering vector
spaces over the field of integers modulo 2 and 4, leading to real and
complex packings, respectively.
• p. 3, last line, the normalization {1/√2A} is incorrect, the factor 1/
√2
apply only to the last four codewords of A and not the first two ones.
Publication VIII
• Selection of codewords is not clearly specified in the paper, we provide
details here:
– In Fig. 1, as we were comparing coding in different manifolds, the
codeword selection was performed using the optimal selection that cor-
respond to maximize the information rate. It follows that the state-
ment p. 5 “When using this scheme, we have selected the codeword
minimizing the flag distance dp over all combinations of Grassman-
nian and orthogonalization codewords.” is erroneous.
– In Fig. 2, the codewords selection was performed with the permutation-
invariant flag chordal distance dp.
• Additionally, there are typos on the labelling of Fig. 2: for Nt = 4, the
curves corresponds to 1, 2, 3, and 5 feedback bits.
68
Errata
69
Errata
70
9HSTFMG*afejbj+
ISBN 978-952-60-5491-9 ISBN 978-952-60-5492-6 (pdf) ISSN-L 1799-4934 ISSN 1799-4934 ISSN 1799-4942 (pdf) Aalto University School of Electrical Engineering Department of Communications and Networking www.aalto.fi
BUSINESS + ECONOMY ART + DESIGN + ARCHITECTURE SCIENCE + TECHNOLOGY CROSSOVER DOCTORAL DISSERTATIONS
Aalto-D
D 211
/2013
Renaud-A
lexandre Pitaval
Coding on F
lag Manifolds for L
imited F
eedback MIM
O System
s A
alto U
nive
rsity
Department of Communications and Networking
Coding on Flag Manifolds for Limited Feedback MIMO Systems
Renaud-Alexandre Pitaval
DOCTORAL DISSERTATIONS